fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/spinlock.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/swap.h>
  10 #include <linux/writeback.h>
  11 #include <linux/pagevec.h>
  12 #include <linux/prefetch.h>
  13 #include <linux/cleancache.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "ctree.h"
  17 #include "btrfs_inode.h"
  18 #include "volumes.h"
  19 #include "check-integrity.h"
  20 #include "locking.h"
  21 #include "rcu-string.h"
  22 #include "backref.h"
  23
  24 static struct kmem_cache *extent_state_cache;
  25 static struct kmem_cache *extent_buffer_cache;
  26 static struct bio_set *btrfs_bioset;
  27
  28 static inline bool extent_state_in_tree(const struct extent_state *state)
  29 {
  30         return !RB_EMPTY_NODE(&state->rb_node);
  31 }
  32
  33 #ifdef CONFIG_BTRFS_DEBUG
  34 static LIST_HEAD(buffers);
  35 static LIST_HEAD(states);
  36
  37 static DEFINE_SPINLOCK(leak_lock);
  38
  39 static inline
  40 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  41 {
  42         unsigned long flags;
  43
  44         spin_lock_irqsave(&leak_lock, flags);
  45         list_add(new, head);
  46         spin_unlock_irqrestore(&leak_lock, flags);
  47 }
  48
  49 static inline
  50 void btrfs_leak_debug_del(struct list_head *entry)
  51 {
  52         unsigned long flags;
  53
  54         spin_lock_irqsave(&leak_lock, flags);
  55         list_del(entry);
  56         spin_unlock_irqrestore(&leak_lock, flags);
  57 }
  58
  59 static inline
  60 void btrfs_leak_debug_check(void)
  61 {
  62         struct extent_state *state;
  63         struct extent_buffer *eb;
  64
  65         while (!list_empty(&states)) {
  66                 state = list_entry(states.next, struct extent_state, leak_list);
  67                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  68                        state->start, state->end, state->state,
  69                        extent_state_in_tree(state),
  70                        atomic_read(&state->refs));
  71                 list_del(&state->leak_list);
  72                 kmem_cache_free(extent_state_cache, state);
  73         }
  74
  75         while (!list_empty(&buffers)) {
  76                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  77                 printk(KERN_ERR "BTRFS: buffer leak start %llu len %lu "
  78                        "refs %d\n",
  79                        eb->start, eb->len, atomic_read(&eb->refs));
  80                 list_del(&eb->leak_list);
  81                 kmem_cache_free(extent_buffer_cache, eb);
  82         }
  83 }
  84
  85 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  86         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  87 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  88                 struct extent_io_tree *tree, u64 start, u64 end)
  89 {
  90         struct inode *inode;
  91         u64 isize;
  92
  93         if (!tree->mapping)
  94                 return;
  95
  96         inode = tree->mapping->host;
  97         isize = i_size_read(inode);
  98         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
  99                 printk_ratelimited(KERN_DEBUG
 100                     "BTRFS: %s: ino %llu isize %llu odd range [%llu,%llu]\n",
 101                                 caller, btrfs_ino(inode), isize, start, end);
 102         }
 103 }
 104 #else
 105 #define btrfs_leak_debug_add(new, head) do {} while (0)
 106 #define btrfs_leak_debug_del(entry)     do {} while (0)
 107 #define btrfs_leak_debug_check()        do {} while (0)
 108 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 109 #endif
 110
 111 #define BUFFER_LRU_MAX 64
 112
 113 struct tree_entry {
 114         u64 start;
 115         u64 end;
 116         struct rb_node rb_node;
 117 };
 118
 119 struct extent_page_data {
 120         struct bio *bio;
 121         struct extent_io_tree *tree;
 122         get_extent_t *get_extent;
 123         unsigned long bio_flags;
 124
 125         /* tells writepage not to lock the state bits for this range
 126          * it still does the unlocking
 127          */
 128         unsigned int extent_locked:1;
 129
 130         /* tells the submit_bio code to use a WRITE_SYNC */
 131         unsigned int sync_io:1;
 132 };
 133
 134 static noinline void flush_write_bio(void *data);
 135 static inline struct btrfs_fs_info *
 136 tree_fs_info(struct extent_io_tree *tree)
 137 {
 138         if (!tree->mapping)
 139                 return NULL;
 140         return btrfs_sb(tree->mapping->host->i_sb);
 141 }
 142
 143 int __init extent_io_init(void)
 144 {
 145         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 146                         sizeof(struct extent_state), 0,
 147                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 148         if (!extent_state_cache)
 149                 return -ENOMEM;
 150
 151         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 152                         sizeof(struct extent_buffer), 0,
 153                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 154         if (!extent_buffer_cache)
 155                 goto free_state_cache;
 156
 157         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 158                                      offsetof(struct btrfs_io_bio, bio));
 159         if (!btrfs_bioset)
 160                 goto free_buffer_cache;
 161
 162         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 163                 goto free_bioset;
 164
 165         return 0;
 166
 167 free_bioset:
 168         bioset_free(btrfs_bioset);
 169         btrfs_bioset = NULL;
 170
 171 free_buffer_cache:
 172         kmem_cache_destroy(extent_buffer_cache);
 173         extent_buffer_cache = NULL;
 174
 175 free_state_cache:
 176         kmem_cache_destroy(extent_state_cache);
 177         extent_state_cache = NULL;
 178         return -ENOMEM;
 179 }
 180
 181 void extent_io_exit(void)
 182 {
 183         btrfs_leak_debug_check();
 184
 185         /*
 186          * Make sure all delayed rcu free are flushed before we
 187          * destroy caches.
 188          */
 189         rcu_barrier();
 190         if (extent_state_cache)
 191                 kmem_cache_destroy(extent_state_cache);
 192         if (extent_buffer_cache)
 193                 kmem_cache_destroy(extent_buffer_cache);
 194         if (btrfs_bioset)
 195                 bioset_free(btrfs_bioset);
 196 }
 197
 198 void extent_io_tree_init(struct extent_io_tree *tree,
 199                          struct address_space *mapping)
 200 {
 201         tree->state = RB_ROOT;
 202         tree->ops = NULL;
 203         tree->dirty_bytes = 0;
 204         spin_lock_init(&tree->lock);
 205         tree->mapping = mapping;
 206 }
 207
 208 static struct extent_state *alloc_extent_state(gfp_t mask)
 209 {
 210         struct extent_state *state;
 211
 212         state = kmem_cache_alloc(extent_state_cache, mask);
 213         if (!state)
 214                 return state;
 215         state->state = 0;
 216         state->private = 0;
 217         RB_CLEAR_NODE(&state->rb_node);
 218         btrfs_leak_debug_add(&state->leak_list, &states);
 219         atomic_set(&state->refs, 1);
 220         init_waitqueue_head(&state->wq);
 221         trace_alloc_extent_state(state, mask, _RET_IP_);
 222         return state;
 223 }
 224
 225 void free_extent_state(struct extent_state *state)
 226 {
 227         if (!state)
 228                 return;
 229         if (atomic_dec_and_test(&state->refs)) {
 230                 WARN_ON(extent_state_in_tree(state));
 231                 btrfs_leak_debug_del(&state->leak_list);
 232                 trace_free_extent_state(state, _RET_IP_);
 233                 kmem_cache_free(extent_state_cache, state);
 234         }
 235 }
 236
 237 static struct rb_node *tree_insert(struct rb_root *root,
 238                                    struct rb_node *search_start,
 239                                    u64 offset,
 240                                    struct rb_node *node,
 241                                    struct rb_node ***p_in,
 242                                    struct rb_node **parent_in)
 243 {
 244         struct rb_node **p;
 245         struct rb_node *parent = NULL;
 246         struct tree_entry *entry;
 247
 248         if (p_in && parent_in) {
 249                 p = *p_in;
 250                 parent = *parent_in;
 251                 goto do_insert;
 252         }
 253
 254         p = search_start ? &search_start : &root->rb_node;
 255         while (*p) {
 256                 parent = *p;
 257                 entry = rb_entry(parent, struct tree_entry, rb_node);
 258
 259                 if (offset < entry->start)
 260                         p = &(*p)->rb_left;
 261                 else if (offset > entry->end)
 262                         p = &(*p)->rb_right;
 263                 else
 264                         return parent;
 265         }
 266
 267 do_insert:
 268         rb_link_node(node, parent, p);
 269         rb_insert_color(node, root);
 270         return NULL;
 271 }
 272
 273 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 274                                       struct rb_node **prev_ret,
 275                                       struct rb_node **next_ret,
 276                                       struct rb_node ***p_ret,
 277                                       struct rb_node **parent_ret)
 278 {
 279         struct rb_root *root = &tree->state;
 280         struct rb_node **n = &root->rb_node;
 281         struct rb_node *prev = NULL;
 282         struct rb_node *orig_prev = NULL;
 283         struct tree_entry *entry;
 284         struct tree_entry *prev_entry = NULL;
 285
 286         while (*n) {
 287                 prev = *n;
 288                 entry = rb_entry(prev, struct tree_entry, rb_node);
 289                 prev_entry = entry;
 290
 291                 if (offset < entry->start)
 292                         n = &(*n)->rb_left;
 293                 else if (offset > entry->end)
 294                         n = &(*n)->rb_right;
 295                 else
 296                         return *n;
 297         }
 298
 299         if (p_ret)
 300                 *p_ret = n;
 301         if (parent_ret)
 302                 *parent_ret = prev;
 303
 304         if (prev_ret) {
 305                 orig_prev = prev;
 306                 while (prev && offset > prev_entry->end) {
 307                         prev = rb_next(prev);
 308                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 309                 }
 310                 *prev_ret = prev;
 311                 prev = orig_prev;
 312         }
 313
 314         if (next_ret) {
 315                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 316                 while (prev && offset < prev_entry->start) {
 317                         prev = rb_prev(prev);
 318                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 319                 }
 320                 *next_ret = prev;
 321         }
 322         return NULL;
 323 }
 324
 325 static inline struct rb_node *
 326 tree_search_for_insert(struct extent_io_tree *tree,
 327                        u64 offset,
 328                        struct rb_node ***p_ret,
 329                        struct rb_node **parent_ret)
 330 {
 331         struct rb_node *prev = NULL;
 332         struct rb_node *ret;
 333
 334         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 335         if (!ret)
 336                 return prev;
 337         return ret;
 338 }
 339
 340 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 341                                           u64 offset)
 342 {
 343         return tree_search_for_insert(tree, offset, NULL, NULL);
 344 }
 345
 346 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 347                      struct extent_state *other)
 348 {
 349         if (tree->ops && tree->ops->merge_extent_hook)
 350                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 351                                              other);
 352 }
 353
 354 /*
 355  * utility function to look for merge candidates inside a given range.
 356  * Any extents with matching state are merged together into a single
 357  * extent in the tree.  Extents with EXTENT_IO in their state field
 358  * are not merged because the end_io handlers need to be able to do
 359  * operations on them without sleeping (or doing allocations/splits).
 360  *
 361  * This should be called with the tree lock held.
 362  */
 363 static void merge_state(struct extent_io_tree *tree,
 364                         struct extent_state *state)
 365 {
 366         struct extent_state *other;
 367         struct rb_node *other_node;
 368
 369         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 370                 return;
 371
 372         other_node = rb_prev(&state->rb_node);
 373         if (other_node) {
 374                 other = rb_entry(other_node, struct extent_state, rb_node);
 375                 if (other->end == state->start - 1 &&
 376                     other->state == state->state) {
 377                         merge_cb(tree, state, other);
 378                         state->start = other->start;
 379                         rb_erase(&other->rb_node, &tree->state);
 380                         RB_CLEAR_NODE(&other->rb_node);
 381                         free_extent_state(other);
 382                 }
 383         }
 384         other_node = rb_next(&state->rb_node);
 385         if (other_node) {
 386                 other = rb_entry(other_node, struct extent_state, rb_node);
 387                 if (other->start == state->end + 1 &&
 388                     other->state == state->state) {
 389                         merge_cb(tree, state, other);
 390                         state->end = other->end;
 391                         rb_erase(&other->rb_node, &tree->state);
 392                         RB_CLEAR_NODE(&other->rb_node);
 393                         free_extent_state(other);
 394                 }
 395         }
 396 }
 397
 398 static void set_state_cb(struct extent_io_tree *tree,
 399                          struct extent_state *state, unsigned *bits)
 400 {
 401         if (tree->ops && tree->ops->set_bit_hook)
 402                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 403 }
 404
 405 static void clear_state_cb(struct extent_io_tree *tree,
 406                            struct extent_state *state, unsigned *bits)
 407 {
 408         if (tree->ops && tree->ops->clear_bit_hook)
 409                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 410 }
 411
 412 static void set_state_bits(struct extent_io_tree *tree,
 413                            struct extent_state *state, unsigned *bits);
 414
 415 /*
 416  * insert an extent_state struct into the tree.  'bits' are set on the
 417  * struct before it is inserted.
 418  *
 419  * This may return -EEXIST if the extent is already there, in which case the
 420  * state struct is freed.
 421  *
 422  * The tree lock is not taken internally.  This is a utility function and
 423  * probably isn't what you want to call (see set/clear_extent_bit).
 424  */
 425 static int insert_state(struct extent_io_tree *tree,
 426                         struct extent_state *state, u64 start, u64 end,
 427                         struct rb_node ***p,
 428                         struct rb_node **parent,
 429                         unsigned *bits)
 430 {
 431         struct rb_node *node;
 432
 433         if (end < start)
 434                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 435                        end, start);
 436         state->start = start;
 437         state->end = end;
 438
 439         set_state_bits(tree, state, bits);
 440
 441         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 442         if (node) {
 443                 struct extent_state *found;
 444                 found = rb_entry(node, struct extent_state, rb_node);
 445                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
 446                        "%llu %llu\n",
 447                        found->start, found->end, start, end);
 448                 return -EEXIST;
 449         }
 450         merge_state(tree, state);
 451         return 0;
 452 }
 453
 454 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 455                      u64 split)
 456 {
 457         if (tree->ops && tree->ops->split_extent_hook)
 458                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 459 }
 460
 461 /*
 462  * split a given extent state struct in two, inserting the preallocated
 463  * struct 'prealloc' as the newly created second half.  'split' indicates an
 464  * offset inside 'orig' where it should be split.
 465  *
 466  * Before calling,
 467  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 468  * are two extent state structs in the tree:
 469  * prealloc: [orig->start, split - 1]
 470  * orig: [ split, orig->end ]
 471  *
 472  * The tree locks are not taken by this function. They need to be held
 473  * by the caller.
 474  */
 475 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 476                        struct extent_state *prealloc, u64 split)
 477 {
 478         struct rb_node *node;
 479
 480         split_cb(tree, orig, split);
 481
 482         prealloc->start = orig->start;
 483         prealloc->end = split - 1;
 484         prealloc->state = orig->state;
 485         orig->start = split;
 486
 487         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 488                            &prealloc->rb_node, NULL, NULL);
 489         if (node) {
 490                 free_extent_state(prealloc);
 491                 return -EEXIST;
 492         }
 493         return 0;
 494 }
 495
 496 static struct extent_state *next_state(struct extent_state *state)
 497 {
 498         struct rb_node *next = rb_next(&state->rb_node);
 499         if (next)
 500                 return rb_entry(next, struct extent_state, rb_node);
 501         else
 502                 return NULL;
 503 }
 504
 505 /*
 506  * utility function to clear some bits in an extent state struct.
 507  * it will optionally wake up any one waiting on this state (wake == 1).
 508  *
 509  * If no bits are set on the state struct after clearing things, the
 510  * struct is freed and removed from the tree
 511  */
 512 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 513                                             struct extent_state *state,
 514                                             unsigned *bits, int wake)
 515 {
 516         struct extent_state *next;
 517         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 518
 519         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 520                 u64 range = state->end - state->start + 1;
 521                 WARN_ON(range > tree->dirty_bytes);
 522                 tree->dirty_bytes -= range;
 523         }
 524         clear_state_cb(tree, state, bits);
 525         state->state &= ~bits_to_clear;
 526         if (wake)
 527                 wake_up(&state->wq);
 528         if (state->state == 0) {
 529                 next = next_state(state);
 530                 if (extent_state_in_tree(state)) {
 531                         rb_erase(&state->rb_node, &tree->state);
 532                         RB_CLEAR_NODE(&state->rb_node);
 533                         free_extent_state(state);
 534                 } else {
 535                         WARN_ON(1);
 536                 }
 537         } else {
 538                 merge_state(tree, state);
 539                 next = next_state(state);
 540         }
 541         return next;
 542 }
 543
 544 static struct extent_state *
 545 alloc_extent_state_atomic(struct extent_state *prealloc)
 546 {
 547         if (!prealloc)
 548                 prealloc = alloc_extent_state(GFP_ATOMIC);
 549
 550         return prealloc;
 551 }
 552
 553 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 554 {
 555         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 556                     "Extent tree was modified by another "
 557                     "thread while locked.");
 558 }
 559
 560 /*
 561  * clear some bits on a range in the tree.  This may require splitting
 562  * or inserting elements in the tree, so the gfp mask is used to
 563  * indicate which allocations or sleeping are allowed.
 564  *
 565  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 566  * the given range from the tree regardless of state (ie for truncate).
 567  *
 568  * the range [start, end] is inclusive.
 569  *
 570  * This takes the tree lock, and returns 0 on success and < 0 on error.
 571  */
 572 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 573                      unsigned bits, int wake, int delete,
 574                      struct extent_state **cached_state,
 575                      gfp_t mask)
 576 {
 577         struct extent_state *state;
 578         struct extent_state *cached;
 579         struct extent_state *prealloc = NULL;
 580         struct rb_node *node;
 581         u64 last_end;
 582         int err;
 583         int clear = 0;
 584
 585         btrfs_debug_check_extent_io_range(tree, start, end);
 586
 587         if (bits & EXTENT_DELALLOC)
 588                 bits |= EXTENT_NORESERVE;
 589
 590         if (delete)
 591                 bits |= ~EXTENT_CTLBITS;
 592         bits |= EXTENT_FIRST_DELALLOC;
 593
 594         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 595                 clear = 1;
 596 again:
 597         if (!prealloc && (mask & __GFP_WAIT)) {
 598                 /*
 599                  * Don't care for allocation failure here because we might end
 600                  * up not needing the pre-allocated extent state at all, which
 601                  * is the case if we only have in the tree extent states that
 602                  * cover our input range and don't cover too any other range.
 603                  * If we end up needing a new extent state we allocate it later.
 604                  */
 605                 prealloc = alloc_extent_state(mask);
 606         }
 607
 608         spin_lock(&tree->lock);
 609         if (cached_state) {
 610                 cached = *cached_state;
 611
 612                 if (clear) {
 613                         *cached_state = NULL;
 614                         cached_state = NULL;
 615                 }
 616
 617                 if (cached && extent_state_in_tree(cached) &&
 618                     cached->start <= start && cached->end > start) {
 619                         if (clear)
 620                                 atomic_dec(&cached->refs);
 621                         state = cached;
 622                         goto hit_next;
 623                 }
 624                 if (clear)
 625                         free_extent_state(cached);
 626         }
 627         /*
 628          * this search will find the extents that end after
 629          * our range starts
 630          */
 631         node = tree_search(tree, start);
 632         if (!node)
 633                 goto out;
 634         state = rb_entry(node, struct extent_state, rb_node);
 635 hit_next:
 636         if (state->start > end)
 637                 goto out;
 638         WARN_ON(state->end < start);
 639         last_end = state->end;
 640
 641         /* the state doesn't have the wanted bits, go ahead */
 642         if (!(state->state & bits)) {
 643                 state = next_state(state);
 644                 goto next;
 645         }
 646
 647         /*
 648          *     | ---- desired range ---- |
 649          *  | state | or
 650          *  | ------------- state -------------- |
 651          *
 652          * We need to split the extent we found, and may flip
 653          * bits on second half.
 654          *
 655          * If the extent we found extends past our range, we
 656          * just split and search again.  It'll get split again
 657          * the next time though.
 658          *
 659          * If the extent we found is inside our range, we clear
 660          * the desired bit on it.
 661          */
 662
 663         if (state->start < start) {
 664                 prealloc = alloc_extent_state_atomic(prealloc);
 665                 BUG_ON(!prealloc);
 666                 err = split_state(tree, state, prealloc, start);
 667                 if (err)
 668                         extent_io_tree_panic(tree, err);
 669
 670                 prealloc = NULL;
 671                 if (err)
 672                         goto out;
 673                 if (state->end <= end) {
 674                         state = clear_state_bit(tree, state, &bits, wake);
 675                         goto next;
 676                 }
 677                 goto search_again;
 678         }
 679         /*
 680          * | ---- desired range ---- |
 681          *                        | state |
 682          * We need to split the extent, and clear the bit
 683          * on the first half
 684          */
 685         if (state->start <= end && state->end > end) {
 686                 prealloc = alloc_extent_state_atomic(prealloc);
 687                 BUG_ON(!prealloc);
 688                 err = split_state(tree, state, prealloc, end + 1);
 689                 if (err)
 690                         extent_io_tree_panic(tree, err);
 691
 692                 if (wake)
 693                         wake_up(&state->wq);
 694
 695                 clear_state_bit(tree, prealloc, &bits, wake);
 696
 697                 prealloc = NULL;
 698                 goto out;
 699         }
 700
 701         state = clear_state_bit(tree, state, &bits, wake);
 702 next:
 703         if (last_end == (u64)-1)
 704                 goto out;
 705         start = last_end + 1;
 706         if (start <= end && state && !need_resched())
 707                 goto hit_next;
 708         goto search_again;
 709
 710 out:
 711         spin_unlock(&tree->lock);
 712         if (prealloc)
 713                 free_extent_state(prealloc);
 714
 715         return 0;
 716
 717 search_again:
 718         if (start > end)
 719                 goto out;
 720         spin_unlock(&tree->lock);
 721         if (mask & __GFP_WAIT)
 722                 cond_resched();
 723         goto again;
 724 }
 725
 726 static void wait_on_state(struct extent_io_tree *tree,
 727                           struct extent_state *state)
 728                 __releases(tree->lock)
 729                 __acquires(tree->lock)
 730 {
 731         DEFINE_WAIT(wait);
 732         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 733         spin_unlock(&tree->lock);
 734         schedule();
 735         spin_lock(&tree->lock);
 736         finish_wait(&state->wq, &wait);
 737 }
 738
 739 /*
 740  * waits for one or more bits to clear on a range in the state tree.
 741  * The range [start, end] is inclusive.
 742  * The tree lock is taken by this function
 743  */
 744 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 745                             unsigned long bits)
 746 {
 747         struct extent_state *state;
 748         struct rb_node *node;
 749
 750         btrfs_debug_check_extent_io_range(tree, start, end);
 751
 752         spin_lock(&tree->lock);
 753 again:
 754         while (1) {
 755                 /*
 756                  * this search will find all the extents that end after
 757                  * our range starts
 758                  */
 759                 node = tree_search(tree, start);
 760 process_node:
 761                 if (!node)
 762                         break;
 763
 764                 state = rb_entry(node, struct extent_state, rb_node);
 765
 766                 if (state->start > end)
 767                         goto out;
 768
 769                 if (state->state & bits) {
 770                         start = state->start;
 771                         atomic_inc(&state->refs);
 772                         wait_on_state(tree, state);
 773                         free_extent_state(state);
 774                         goto again;
 775                 }
 776                 start = state->end + 1;
 777
 778                 if (start > end)
 779                         break;
 780
 781                 if (!cond_resched_lock(&tree->lock)) {
 782                         node = rb_next(node);
 783                         goto process_node;
 784                 }
 785         }
 786 out:
 787         spin_unlock(&tree->lock);
 788 }
 789
 790 static void set_state_bits(struct extent_io_tree *tree,
 791                            struct extent_state *state,
 792                            unsigned *bits)
 793 {
 794         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 795
 796         set_state_cb(tree, state, bits);
 797         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 798                 u64 range = state->end - state->start + 1;
 799                 tree->dirty_bytes += range;
 800         }
 801         state->state |= bits_to_set;
 802 }
 803
 804 static void cache_state_if_flags(struct extent_state *state,
 805                                  struct extent_state **cached_ptr,
 806                                  unsigned flags)
 807 {
 808         if (cached_ptr && !(*cached_ptr)) {
 809                 if (!flags || (state->state & flags)) {
 810                         *cached_ptr = state;
 811                         atomic_inc(&state->refs);
 812                 }
 813         }
 814 }
 815
 816 static void cache_state(struct extent_state *state,
 817                         struct extent_state **cached_ptr)
 818 {
 819         return cache_state_if_flags(state, cached_ptr,
 820                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 821 }
 822
 823 /*
 824  * set some bits on a range in the tree.  This may require allocations or
 825  * sleeping, so the gfp mask is used to indicate what is allowed.
 826  *
 827  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 828  * part of the range already has the desired bits set.  The start of the
 829  * existing range is returned in failed_start in this case.
 830  *
 831  * [start, end] is inclusive This takes the tree lock.
 832  */
 833
 834 static int __must_check
 835 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 836                  unsigned bits, unsigned exclusive_bits,
 837                  u64 *failed_start, struct extent_state **cached_state,
 838                  gfp_t mask)
 839 {
 840         struct extent_state *state;
 841         struct extent_state *prealloc = NULL;
 842         struct rb_node *node;
 843         struct rb_node **p;
 844         struct rb_node *parent;
 845         int err = 0;
 846         u64 last_start;
 847         u64 last_end;
 848
 849         btrfs_debug_check_extent_io_range(tree, start, end);
 850
 851         bits |= EXTENT_FIRST_DELALLOC;
 852 again:
 853         if (!prealloc && (mask & __GFP_WAIT)) {
 854                 prealloc = alloc_extent_state(mask);
 855                 BUG_ON(!prealloc);
 856         }
 857
 858         spin_lock(&tree->lock);
 859         if (cached_state && *cached_state) {
 860                 state = *cached_state;
 861                 if (state->start <= start && state->end > start &&
 862                     extent_state_in_tree(state)) {
 863                         node = &state->rb_node;
 864                         goto hit_next;
 865                 }
 866         }
 867         /*
 868          * this search will find all the extents that end after
 869          * our range starts.
 870          */
 871         node = tree_search_for_insert(tree, start, &p, &parent);
 872         if (!node) {
 873                 prealloc = alloc_extent_state_atomic(prealloc);
 874                 BUG_ON(!prealloc);
 875                 err = insert_state(tree, prealloc, start, end,
 876                                    &p, &parent, &bits);
 877                 if (err)
 878                         extent_io_tree_panic(tree, err);
 879
 880                 cache_state(prealloc, cached_state);
 881                 prealloc = NULL;
 882                 goto out;
 883         }
 884         state = rb_entry(node, struct extent_state, rb_node);
 885 hit_next:
 886         last_start = state->start;
 887         last_end = state->end;
 888
 889         /*
 890          * | ---- desired range ---- |
 891          * | state |
 892          *
 893          * Just lock what we found and keep going
 894          */
 895         if (state->start == start && state->end <= end) {
 896                 if (state->state & exclusive_bits) {
 897                         *failed_start = state->start;
 898                         err = -EEXIST;
 899                         goto out;
 900                 }
 901
 902                 set_state_bits(tree, state, &bits);
 903                 cache_state(state, cached_state);
 904                 merge_state(tree, state);
 905                 if (last_end == (u64)-1)
 906                         goto out;
 907                 start = last_end + 1;
 908                 state = next_state(state);
 909                 if (start < end && state && state->start == start &&
 910                     !need_resched())
 911                         goto hit_next;
 912                 goto search_again;
 913         }
 914
 915         /*
 916          *     | ---- desired range ---- |
 917          * | state |
 918          *   or
 919          * | ------------- state -------------- |
 920          *
 921          * We need to split the extent we found, and may flip bits on
 922          * second half.
 923          *
 924          * If the extent we found extends past our
 925          * range, we just split and search again.  It'll get split
 926          * again the next time though.
 927          *
 928          * If the extent we found is inside our range, we set the
 929          * desired bit on it.
 930          */
 931         if (state->start < start) {
 932                 if (state->state & exclusive_bits) {
 933                         *failed_start = start;
 934                         err = -EEXIST;
 935                         goto out;
 936                 }
 937
 938                 prealloc = alloc_extent_state_atomic(prealloc);
 939                 BUG_ON(!prealloc);
 940                 err = split_state(tree, state, prealloc, start);
 941                 if (err)
 942                         extent_io_tree_panic(tree, err);
 943
 944                 prealloc = NULL;
 945                 if (err)
 946                         goto out;
 947                 if (state->end <= end) {
 948                         set_state_bits(tree, state, &bits);
 949                         cache_state(state, cached_state);
 950                         merge_state(tree, state);
 951                         if (last_end == (u64)-1)
 952                                 goto out;
 953                         start = last_end + 1;
 954                         state = next_state(state);
 955                         if (start < end && state && state->start == start &&
 956                             !need_resched())
 957                                 goto hit_next;
 958                 }
 959                 goto search_again;
 960         }
 961         /*
 962          * | ---- desired range ---- |
 963          *     | state | or               | state |
 964          *
 965          * There's a hole, we need to insert something in it and
 966          * ignore the extent we found.
 967          */
 968         if (state->start > start) {
 969                 u64 this_end;
 970                 if (end < last_start)
 971                         this_end = end;
 972                 else
 973                         this_end = last_start - 1;
 974
 975                 prealloc = alloc_extent_state_atomic(prealloc);
 976                 BUG_ON(!prealloc);
 977
 978                 /*
 979                  * Avoid to free 'prealloc' if it can be merged with
 980                  * the later extent.
 981                  */
 982                 err = insert_state(tree, prealloc, start, this_end,
 983                                    NULL, NULL, &bits);
 984                 if (err)
 985                         extent_io_tree_panic(tree, err);
 986
 987                 cache_state(prealloc, cached_state);
 988                 prealloc = NULL;
 989                 start = this_end + 1;
 990                 goto search_again;
 991         }
 992         /*
 993          * | ---- desired range ---- |
 994          *                        | state |
 995          * We need to split the extent, and set the bit
 996          * on the first half
 997          */
 998         if (state->start <= end && state->end > end) {
 999                 if (state->state & exclusive_bits) {
1000                         *failed_start = start;
1001                         err = -EEXIST;
1002                         goto out;
1003                 }
1004
1005                 prealloc = alloc_extent_state_atomic(prealloc);
1006                 BUG_ON(!prealloc);
1007                 err = split_state(tree, state, prealloc, end + 1);
1008                 if (err)
1009                         extent_io_tree_panic(tree, err);
1010
1011                 set_state_bits(tree, prealloc, &bits);
1012                 cache_state(prealloc, cached_state);
1013                 merge_state(tree, prealloc);
1014                 prealloc = NULL;
1015                 goto out;
1016         }
1017
1018         goto search_again;
1019
1020 out:
1021         spin_unlock(&tree->lock);
1022         if (prealloc)
1023                 free_extent_state(prealloc);
1024
1025         return err;
1026
1027 search_again:
1028         if (start > end)
1029                 goto out;
1030         spin_unlock(&tree->lock);
1031         if (mask & __GFP_WAIT)
1032                 cond_resched();
1033         goto again;
1034 }
1035
1036 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1037                    unsigned bits, u64 * failed_start,
1038                    struct extent_state **cached_state, gfp_t mask)
1039 {
1040         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1041                                 cached_state, mask);
1042 }
1043
1044
1045 /**
1046  * convert_extent_bit - convert all bits in a given range from one bit to
1047  *                      another
1048  * @tree:       the io tree to search
1049  * @start:      the start offset in bytes
1050  * @end:        the end offset in bytes (inclusive)
1051  * @bits:       the bits to set in this range
1052  * @clear_bits: the bits to clear in this range
1053  * @cached_state:       state that we're going to cache
1054  * @mask:       the allocation mask
1055  *
1056  * This will go through and set bits for the given range.  If any states exist
1057  * already in this range they are set with the given bit and cleared of the
1058  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1059  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1060  * boundary bits like LOCK.
1061  */
1062 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1063                        unsigned bits, unsigned clear_bits,
1064                        struct extent_state **cached_state, gfp_t mask)
1065 {
1066         struct extent_state *state;
1067         struct extent_state *prealloc = NULL;
1068         struct rb_node *node;
1069         struct rb_node **p;
1070         struct rb_node *parent;
1071         int err = 0;
1072         u64 last_start;
1073         u64 last_end;
1074         bool first_iteration = true;
1075
1076         btrfs_debug_check_extent_io_range(tree, start, end);
1077
1078 again:
1079         if (!prealloc && (mask & __GFP_WAIT)) {
1080                 /*
1081                  * Best effort, don't worry if extent state allocation fails
1082                  * here for the first iteration. We might have a cached state
1083                  * that matches exactly the target range, in which case no
1084                  * extent state allocations are needed. We'll only know this
1085                  * after locking the tree.
1086                  */
1087                 prealloc = alloc_extent_state(mask);
1088                 if (!prealloc && !first_iteration)
1089                         return -ENOMEM;
1090         }
1091
1092         spin_lock(&tree->lock);
1093         if (cached_state && *cached_state) {
1094                 state = *cached_state;
1095                 if (state->start <= start && state->end > start &&
1096                     extent_state_in_tree(state)) {
1097                         node = &state->rb_node;
1098                         goto hit_next;
1099                 }
1100         }
1101
1102         /*
1103          * this search will find all the extents that end after
1104          * our range starts.
1105          */
1106         node = tree_search_for_insert(tree, start, &p, &parent);
1107         if (!node) {
1108                 prealloc = alloc_extent_state_atomic(prealloc);
1109                 if (!prealloc) {
1110                         err = -ENOMEM;
1111                         goto out;
1112                 }
1113                 err = insert_state(tree, prealloc, start, end,
1114                                    &p, &parent, &bits);
1115                 if (err)
1116                         extent_io_tree_panic(tree, err);
1117                 cache_state(prealloc, cached_state);
1118                 prealloc = NULL;
1119                 goto out;
1120         }
1121         state = rb_entry(node, struct extent_state, rb_node);
1122 hit_next:
1123         last_start = state->start;
1124         last_end = state->end;
1125
1126         /*
1127          * | ---- desired range ---- |
1128          * | state |
1129          *
1130          * Just lock what we found and keep going
1131          */
1132         if (state->start == start && state->end <= end) {
1133                 set_state_bits(tree, state, &bits);
1134                 cache_state(state, cached_state);
1135                 state = clear_state_bit(tree, state, &clear_bits, 0);
1136                 if (last_end == (u64)-1)
1137                         goto out;
1138                 start = last_end + 1;
1139                 if (start < end && state && state->start == start &&
1140                     !need_resched())
1141                         goto hit_next;
1142                 goto search_again;
1143         }
1144
1145         /*
1146          *     | ---- desired range ---- |
1147          * | state |
1148          *   or
1149          * | ------------- state -------------- |
1150          *
1151          * We need to split the extent we found, and may flip bits on
1152          * second half.
1153          *
1154          * If the extent we found extends past our
1155          * range, we just split and search again.  It'll get split
1156          * again the next time though.
1157          *
1158          * If the extent we found is inside our range, we set the
1159          * desired bit on it.
1160          */
1161         if (state->start < start) {
1162                 prealloc = alloc_extent_state_atomic(prealloc);
1163                 if (!prealloc) {
1164                         err = -ENOMEM;
1165                         goto out;
1166                 }
1167                 err = split_state(tree, state, prealloc, start);
1168                 if (err)
1169                         extent_io_tree_panic(tree, err);
1170                 prealloc = NULL;
1171                 if (err)
1172                         goto out;
1173                 if (state->end <= end) {
1174                         set_state_bits(tree, state, &bits);
1175                         cache_state(state, cached_state);
1176                         state = clear_state_bit(tree, state, &clear_bits, 0);
1177                         if (last_end == (u64)-1)
1178                                 goto out;
1179                         start = last_end + 1;
1180                         if (start < end && state && state->start == start &&
1181                             !need_resched())
1182                                 goto hit_next;
1183                 }
1184                 goto search_again;
1185         }
1186         /*
1187          * | ---- desired range ---- |
1188          *     | state | or               | state |
1189          *
1190          * There's a hole, we need to insert something in it and
1191          * ignore the extent we found.
1192          */
1193         if (state->start > start) {
1194                 u64 this_end;
1195                 if (end < last_start)
1196                         this_end = end;
1197                 else
1198                         this_end = last_start - 1;
1199
1200                 prealloc = alloc_extent_state_atomic(prealloc);
1201                 if (!prealloc) {
1202                         err = -ENOMEM;
1203                         goto out;
1204                 }
1205
1206                 /*
1207                  * Avoid to free 'prealloc' if it can be merged with
1208                  * the later extent.
1209                  */
1210                 err = insert_state(tree, prealloc, start, this_end,
1211                                    NULL, NULL, &bits);
1212                 if (err)
1213                         extent_io_tree_panic(tree, err);
1214                 cache_state(prealloc, cached_state);
1215                 prealloc = NULL;
1216                 start = this_end + 1;
1217                 goto search_again;
1218         }
1219         /*
1220          * | ---- desired range ---- |
1221          *                        | state |
1222          * We need to split the extent, and set the bit
1223          * on the first half
1224          */
1225         if (state->start <= end && state->end > end) {
1226                 prealloc = alloc_extent_state_atomic(prealloc);
1227                 if (!prealloc) {
1228                         err = -ENOMEM;
1229                         goto out;
1230                 }
1231
1232                 err = split_state(tree, state, prealloc, end + 1);
1233                 if (err)
1234                         extent_io_tree_panic(tree, err);
1235
1236                 set_state_bits(tree, prealloc, &bits);
1237                 cache_state(prealloc, cached_state);
1238                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1239                 prealloc = NULL;
1240                 goto out;
1241         }
1242
1243         goto search_again;
1244
1245 out:
1246         spin_unlock(&tree->lock);
1247         if (prealloc)
1248                 free_extent_state(prealloc);
1249
1250         return err;
1251
1252 search_again:
1253         if (start > end)
1254                 goto out;
1255         spin_unlock(&tree->lock);
1256         if (mask & __GFP_WAIT)
1257                 cond_resched();
1258         first_iteration = false;
1259         goto again;
1260 }
1261
1262 /* wrappers around set/clear extent bit */
1263 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1264                      gfp_t mask)
1265 {
1266         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1267                               NULL, mask);
1268 }
1269
1270 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1271                     unsigned bits, gfp_t mask)
1272 {
1273         return set_extent_bit(tree, start, end, bits, NULL,
1274                               NULL, mask);
1275 }
1276
1277 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1278                       unsigned bits, gfp_t mask)
1279 {
1280         int wake = 0;
1281
1282         if (bits & EXTENT_LOCKED)
1283                 wake = 1;
1284
1285         return clear_extent_bit(tree, start, end, bits, wake, 0, NULL, mask);
1286 }
1287
1288 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1289                         struct extent_state **cached_state, gfp_t mask)
1290 {
1291         return set_extent_bit(tree, start, end,
1292                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1293                               NULL, cached_state, mask);
1294 }
1295
1296 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1297                       struct extent_state **cached_state, gfp_t mask)
1298 {
1299         return set_extent_bit(tree, start, end,
1300                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1301                               NULL, cached_state, mask);
1302 }
1303
1304 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1305                        gfp_t mask)
1306 {
1307         return clear_extent_bit(tree, start, end,
1308                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1309                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1310 }
1311
1312 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1313                      gfp_t mask)
1314 {
1315         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1316                               NULL, mask);
1317 }
1318
1319 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1320                         struct extent_state **cached_state, gfp_t mask)
1321 {
1322         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1323                               cached_state, mask);
1324 }
1325
1326 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1327                           struct extent_state **cached_state, gfp_t mask)
1328 {
1329         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1330                                 cached_state, mask);
1331 }
1332
1333 /*
1334  * either insert or lock state struct between start and end use mask to tell
1335  * us if waiting is desired.
1336  */
1337 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1338                      unsigned bits, struct extent_state **cached_state)
1339 {
1340         int err;
1341         u64 failed_start;
1342
1343         while (1) {
1344                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1345                                        EXTENT_LOCKED, &failed_start,
1346                                        cached_state, GFP_NOFS);
1347                 if (err == -EEXIST) {
1348                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1349                         start = failed_start;
1350                 } else
1351                         break;
1352                 WARN_ON(start > end);
1353         }
1354         return err;
1355 }
1356
1357 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1358 {
1359         return lock_extent_bits(tree, start, end, 0, NULL);
1360 }
1361
1362 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1363 {
1364         int err;
1365         u64 failed_start;
1366
1367         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1368                                &failed_start, NULL, GFP_NOFS);
1369         if (err == -EEXIST) {
1370                 if (failed_start > start)
1371                         clear_extent_bit(tree, start, failed_start - 1,
1372                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1373                 return 0;
1374         }
1375         return 1;
1376 }
1377
1378 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1379                          struct extent_state **cached, gfp_t mask)
1380 {
1381         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1382                                 mask);
1383 }
1384
1385 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1386 {
1387         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1388                                 GFP_NOFS);
1389 }
1390
1391 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1392 {
1393         unsigned long index = start >> PAGE_CACHE_SHIFT;
1394         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1395         struct page *page;
1396
1397         while (index <= end_index) {
1398                 page = find_get_page(inode->i_mapping, index);
1399                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1400                 clear_page_dirty_for_io(page);
1401                 page_cache_release(page);
1402                 index++;
1403         }
1404         return 0;
1405 }
1406
1407 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1408 {
1409         unsigned long index = start >> PAGE_CACHE_SHIFT;
1410         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1411         struct page *page;
1412
1413         while (index <= end_index) {
1414                 page = find_get_page(inode->i_mapping, index);
1415                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1416                 __set_page_dirty_nobuffers(page);
1417                 account_page_redirty(page);
1418                 page_cache_release(page);
1419                 index++;
1420         }
1421         return 0;
1422 }
1423
1424 /*
1425  * helper function to set both pages and extents in the tree writeback
1426  */
1427 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1428 {
1429         unsigned long index = start >> PAGE_CACHE_SHIFT;
1430         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1431         struct page *page;
1432
1433         while (index <= end_index) {
1434                 page = find_get_page(tree->mapping, index);
1435                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1436                 set_page_writeback(page);
1437                 page_cache_release(page);
1438                 index++;
1439         }
1440         return 0;
1441 }
1442
1443 /* find the first state struct with 'bits' set after 'start', and
1444  * return it.  tree->lock must be held.  NULL will returned if
1445  * nothing was found after 'start'
1446  */
1447 static struct extent_state *
1448 find_first_extent_bit_state(struct extent_io_tree *tree,
1449                             u64 start, unsigned bits)
1450 {
1451         struct rb_node *node;
1452         struct extent_state *state;
1453
1454         /*
1455          * this search will find all the extents that end after
1456          * our range starts.
1457          */
1458         node = tree_search(tree, start);
1459         if (!node)
1460                 goto out;
1461
1462         while (1) {
1463                 state = rb_entry(node, struct extent_state, rb_node);
1464                 if (state->end >= start && (state->state & bits))
1465                         return state;
1466
1467                 node = rb_next(node);
1468                 if (!node)
1469                         break;
1470         }
1471 out:
1472         return NULL;
1473 }
1474
1475 /*
1476  * find the first offset in the io tree with 'bits' set. zero is
1477  * returned if we find something, and *start_ret and *end_ret are
1478  * set to reflect the state struct that was found.
1479  *
1480  * If nothing was found, 1 is returned. If found something, return 0.
1481  */
1482 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1483                           u64 *start_ret, u64 *end_ret, unsigned bits,
1484                           struct extent_state **cached_state)
1485 {
1486         struct extent_state *state;
1487         struct rb_node *n;
1488         int ret = 1;
1489
1490         spin_lock(&tree->lock);
1491         if (cached_state && *cached_state) {
1492                 state = *cached_state;
1493                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1494                         n = rb_next(&state->rb_node);
1495                         while (n) {
1496                                 state = rb_entry(n, struct extent_state,
1497                                                  rb_node);
1498                                 if (state->state & bits)
1499                                         goto got_it;
1500                                 n = rb_next(n);
1501                         }
1502                         free_extent_state(*cached_state);
1503                         *cached_state = NULL;
1504                         goto out;
1505                 }
1506                 free_extent_state(*cached_state);
1507                 *cached_state = NULL;
1508         }
1509
1510         state = find_first_extent_bit_state(tree, start, bits);
1511 got_it:
1512         if (state) {
1513                 cache_state_if_flags(state, cached_state, 0);
1514                 *start_ret = state->start;
1515                 *end_ret = state->end;
1516                 ret = 0;
1517         }
1518 out:
1519         spin_unlock(&tree->lock);
1520         return ret;
1521 }
1522
1523 /*
1524  * find a contiguous range of bytes in the file marked as delalloc, not
1525  * more than 'max_bytes'.  start and end are used to return the range,
1526  *
1527  * 1 is returned if we find something, 0 if nothing was in the tree
1528  */
1529 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1530                                         u64 *start, u64 *end, u64 max_bytes,
1531                                         struct extent_state **cached_state)
1532 {
1533         struct rb_node *node;
1534         struct extent_state *state;
1535         u64 cur_start = *start;
1536         u64 found = 0;
1537         u64 total_bytes = 0;
1538
1539         spin_lock(&tree->lock);
1540
1541         /*
1542          * this search will find all the extents that end after
1543          * our range starts.
1544          */
1545         node = tree_search(tree, cur_start);
1546         if (!node) {
1547                 if (!found)
1548                         *end = (u64)-1;
1549                 goto out;
1550         }
1551
1552         while (1) {
1553                 state = rb_entry(node, struct extent_state, rb_node);
1554                 if (found && (state->start != cur_start ||
1555                               (state->state & EXTENT_BOUNDARY))) {
1556                         goto out;
1557                 }
1558                 if (!(state->state & EXTENT_DELALLOC)) {
1559                         if (!found)
1560                                 *end = state->end;
1561                         goto out;
1562                 }
1563                 if (!found) {
1564                         *start = state->start;
1565                         *cached_state = state;
1566                         atomic_inc(&state->refs);
1567                 }
1568                 found++;
1569                 *end = state->end;
1570                 cur_start = state->end + 1;
1571                 node = rb_next(node);
1572                 total_bytes += state->end - state->start + 1;
1573                 if (total_bytes >= max_bytes)
1574                         break;
1575                 if (!node)
1576                         break;
1577         }
1578 out:
1579         spin_unlock(&tree->lock);
1580         return found;
1581 }
1582
1583 static noinline void __unlock_for_delalloc(struct inode *inode,
1584                                            struct page *locked_page,
1585                                            u64 start, u64 end)
1586 {
1587         int ret;
1588         struct page *pages[16];
1589         unsigned long index = start >> PAGE_CACHE_SHIFT;
1590         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1591         unsigned long nr_pages = end_index - index + 1;
1592         int i;
1593
1594         if (index == locked_page->index && end_index == index)
1595                 return;
1596
1597         while (nr_pages > 0) {
1598                 ret = find_get_pages_contig(inode->i_mapping, index,
1599                                      min_t(unsigned long, nr_pages,
1600                                      ARRAY_SIZE(pages)), pages);
1601                 for (i = 0; i < ret; i++) {
1602                         if (pages[i] != locked_page)
1603                                 unlock_page(pages[i]);
1604                         page_cache_release(pages[i]);
1605                 }
1606                 nr_pages -= ret;
1607                 index += ret;
1608                 cond_resched();
1609         }
1610 }
1611
1612 static noinline int lock_delalloc_pages(struct inode *inode,
1613                                         struct page *locked_page,
1614                                         u64 delalloc_start,
1615                                         u64 delalloc_end)
1616 {
1617         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1618         unsigned long start_index = index;
1619         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1620         unsigned long pages_locked = 0;
1621         struct page *pages[16];
1622         unsigned long nrpages;
1623         int ret;
1624         int i;
1625
1626         /* the caller is responsible for locking the start index */
1627         if (index == locked_page->index && index == end_index)
1628                 return 0;
1629
1630         /* skip the page at the start index */
1631         nrpages = end_index - index + 1;
1632         while (nrpages > 0) {
1633                 ret = find_get_pages_contig(inode->i_mapping, index,
1634                                      min_t(unsigned long,
1635                                      nrpages, ARRAY_SIZE(pages)), pages);
1636                 if (ret == 0) {
1637                         ret = -EAGAIN;
1638                         goto done;
1639                 }
1640                 /* now we have an array of pages, lock them all */
1641                 for (i = 0; i < ret; i++) {
1642                         /*
1643                          * the caller is taking responsibility for
1644                          * locked_page
1645                          */
1646                         if (pages[i] != locked_page) {
1647                                 lock_page(pages[i]);
1648                                 if (!PageDirty(pages[i]) ||
1649                                     pages[i]->mapping != inode->i_mapping) {
1650                                         ret = -EAGAIN;
1651                                         unlock_page(pages[i]);
1652                                         page_cache_release(pages[i]);
1653                                         goto done;
1654                                 }
1655                         }
1656                         page_cache_release(pages[i]);
1657                         pages_locked++;
1658                 }
1659                 nrpages -= ret;
1660                 index += ret;
1661                 cond_resched();
1662         }
1663         ret = 0;
1664 done:
1665         if (ret && pages_locked) {
1666                 __unlock_for_delalloc(inode, locked_page,
1667                               delalloc_start,
1668                               ((u64)(start_index + pages_locked - 1)) <<
1669                               PAGE_CACHE_SHIFT);
1670         }
1671         return ret;
1672 }
1673
1674 /*
1675  * find a contiguous range of bytes in the file marked as delalloc, not
1676  * more than 'max_bytes'.  start and end are used to return the range,
1677  *
1678  * 1 is returned if we find something, 0 if nothing was in the tree
1679  */
1680 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1681                                     struct extent_io_tree *tree,
1682                                     struct page *locked_page, u64 *start,
1683                                     u64 *end, u64 max_bytes)
1684 {
1685         u64 delalloc_start;
1686         u64 delalloc_end;
1687         u64 found;
1688         struct extent_state *cached_state = NULL;
1689         int ret;
1690         int loops = 0;
1691
1692 again:
1693         /* step one, find a bunch of delalloc bytes starting at start */
1694         delalloc_start = *start;
1695         delalloc_end = 0;
1696         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1697                                     max_bytes, &cached_state);
1698         if (!found || delalloc_end <= *start) {
1699                 *start = delalloc_start;
1700                 *end = delalloc_end;
1701                 free_extent_state(cached_state);
1702                 return 0;
1703         }
1704
1705         /*
1706          * start comes from the offset of locked_page.  We have to lock
1707          * pages in order, so we can't process delalloc bytes before
1708          * locked_page
1709          */
1710         if (delalloc_start < *start)
1711                 delalloc_start = *start;
1712
1713         /*
1714          * make sure to limit the number of pages we try to lock down
1715          */
1716         if (delalloc_end + 1 - delalloc_start > max_bytes)
1717                 delalloc_end = delalloc_start + max_bytes - 1;
1718
1719         /* step two, lock all the pages after the page that has start */
1720         ret = lock_delalloc_pages(inode, locked_page,
1721                                   delalloc_start, delalloc_end);
1722         if (ret == -EAGAIN) {
1723                 /* some of the pages are gone, lets avoid looping by
1724                  * shortening the size of the delalloc range we're searching
1725                  */
1726                 free_extent_state(cached_state);
1727                 cached_state = NULL;
1728                 if (!loops) {
1729                         max_bytes = PAGE_CACHE_SIZE;
1730                         loops = 1;
1731                         goto again;
1732                 } else {
1733                         found = 0;
1734                         goto out_failed;
1735                 }
1736         }
1737         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1738
1739         /* step three, lock the state bits for the whole range */
1740         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1741
1742         /* then test to make sure it is all still delalloc */
1743         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1744                              EXTENT_DELALLOC, 1, cached_state);
1745         if (!ret) {
1746                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1747                                      &cached_state, GFP_NOFS);
1748                 __unlock_for_delalloc(inode, locked_page,
1749                               delalloc_start, delalloc_end);
1750                 cond_resched();
1751                 goto again;
1752         }
1753         free_extent_state(cached_state);
1754         *start = delalloc_start;
1755         *end = delalloc_end;
1756 out_failed:
1757         return found;
1758 }
1759
1760 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1761                                  struct page *locked_page,
1762                                  unsigned clear_bits,
1763                                  unsigned long page_ops)
1764 {
1765         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1766         int ret;
1767         struct page *pages[16];
1768         unsigned long index = start >> PAGE_CACHE_SHIFT;
1769         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1770         unsigned long nr_pages = end_index - index + 1;
1771         int i;
1772
1773         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1774         if (page_ops == 0)
1775                 return 0;
1776
1777         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1778                 mapping_set_error(inode->i_mapping, -EIO);
1779
1780         while (nr_pages > 0) {
1781                 ret = find_get_pages_contig(inode->i_mapping, index,
1782                                      min_t(unsigned long,
1783                                      nr_pages, ARRAY_SIZE(pages)), pages);
1784                 for (i = 0; i < ret; i++) {
1785
1786                         if (page_ops & PAGE_SET_PRIVATE2)
1787                                 SetPagePrivate2(pages[i]);
1788
1789                         if (pages[i] == locked_page) {
1790                                 page_cache_release(pages[i]);
1791                                 continue;
1792                         }
1793                         if (page_ops & PAGE_CLEAR_DIRTY)
1794                                 clear_page_dirty_for_io(pages[i]);
1795                         if (page_ops & PAGE_SET_WRITEBACK)
1796                                 set_page_writeback(pages[i]);
1797                         if (page_ops & PAGE_SET_ERROR)
1798                                 SetPageError(pages[i]);
1799                         if (page_ops & PAGE_END_WRITEBACK)
1800                                 end_page_writeback(pages[i]);
1801                         if (page_ops & PAGE_UNLOCK)
1802                                 unlock_page(pages[i]);
1803                         page_cache_release(pages[i]);
1804                 }
1805                 nr_pages -= ret;
1806                 index += ret;
1807                 cond_resched();
1808         }
1809         return 0;
1810 }
1811
1812 /*
1813  * count the number of bytes in the tree that have a given bit(s)
1814  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1815  * cached.  The total number found is returned.
1816  */
1817 u64 count_range_bits(struct extent_io_tree *tree,
1818                      u64 *start, u64 search_end, u64 max_bytes,
1819                      unsigned bits, int contig)
1820 {
1821         struct rb_node *node;
1822         struct extent_state *state;
1823         u64 cur_start = *start;
1824         u64 total_bytes = 0;
1825         u64 last = 0;
1826         int found = 0;
1827
1828         if (WARN_ON(search_end <= cur_start))
1829                 return 0;
1830
1831         spin_lock(&tree->lock);
1832         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1833                 total_bytes = tree->dirty_bytes;
1834                 goto out;
1835         }
1836         /*
1837          * this search will find all the extents that end after
1838          * our range starts.
1839          */
1840         node = tree_search(tree, cur_start);
1841         if (!node)
1842                 goto out;
1843
1844         while (1) {
1845                 state = rb_entry(node, struct extent_state, rb_node);
1846                 if (state->start > search_end)
1847                         break;
1848                 if (contig && found && state->start > last + 1)
1849                         break;
1850                 if (state->end >= cur_start && (state->state & bits) == bits) {
1851                         total_bytes += min(search_end, state->end) + 1 -
1852                                        max(cur_start, state->start);
1853                         if (total_bytes >= max_bytes)
1854                                 break;
1855                         if (!found) {
1856                                 *start = max(cur_start, state->start);
1857                                 found = 1;
1858                         }
1859                         last = state->end;
1860                 } else if (contig && found) {
1861                         break;
1862                 }
1863                 node = rb_next(node);
1864                 if (!node)
1865                         break;
1866         }
1867 out:
1868         spin_unlock(&tree->lock);
1869         return total_bytes;
1870 }
1871
1872 /*
1873  * set the private field for a given byte offset in the tree.  If there isn't
1874  * an extent_state there already, this does nothing.
1875  */
1876 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1877 {
1878         struct rb_node *node;
1879         struct extent_state *state;
1880         int ret = 0;
1881
1882         spin_lock(&tree->lock);
1883         /*
1884          * this search will find all the extents that end after
1885          * our range starts.
1886          */
1887         node = tree_search(tree, start);
1888         if (!node) {
1889                 ret = -ENOENT;
1890                 goto out;
1891         }
1892         state = rb_entry(node, struct extent_state, rb_node);
1893         if (state->start != start) {
1894                 ret = -ENOENT;
1895                 goto out;
1896         }
1897         state->private = private;
1898 out:
1899         spin_unlock(&tree->lock);
1900         return ret;
1901 }
1902
1903 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1904 {
1905         struct rb_node *node;
1906         struct extent_state *state;
1907         int ret = 0;
1908
1909         spin_lock(&tree->lock);
1910         /*
1911          * this search will find all the extents that end after
1912          * our range starts.
1913          */
1914         node = tree_search(tree, start);
1915         if (!node) {
1916                 ret = -ENOENT;
1917                 goto out;
1918         }
1919         state = rb_entry(node, struct extent_state, rb_node);
1920         if (state->start != start) {
1921                 ret = -ENOENT;
1922                 goto out;
1923         }
1924         *private = state->private;
1925 out:
1926         spin_unlock(&tree->lock);
1927         return ret;
1928 }
1929
1930 /*
1931  * searches a range in the state tree for a given mask.
1932  * If 'filled' == 1, this returns 1 only if every extent in the tree
1933  * has the bits set.  Otherwise, 1 is returned if any bit in the
1934  * range is found set.
1935  */
1936 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1937                    unsigned bits, int filled, struct extent_state *cached)
1938 {
1939         struct extent_state *state = NULL;
1940         struct rb_node *node;
1941         int bitset = 0;
1942
1943         spin_lock(&tree->lock);
1944         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1945             cached->end > start)
1946                 node = &cached->rb_node;
1947         else
1948                 node = tree_search(tree, start);
1949         while (node && start <= end) {
1950                 state = rb_entry(node, struct extent_state, rb_node);
1951
1952                 if (filled && state->start > start) {
1953                         bitset = 0;
1954                         break;
1955                 }
1956
1957                 if (state->start > end)
1958                         break;
1959
1960                 if (state->state & bits) {
1961                         bitset = 1;
1962                         if (!filled)
1963                                 break;
1964                 } else if (filled) {
1965                         bitset = 0;
1966                         break;
1967                 }
1968
1969                 if (state->end == (u64)-1)
1970                         break;
1971
1972                 start = state->end + 1;
1973                 if (start > end)
1974                         break;
1975                 node = rb_next(node);
1976                 if (!node) {
1977                         if (filled)
1978                                 bitset = 0;
1979                         break;
1980                 }
1981         }
1982         spin_unlock(&tree->lock);
1983         return bitset;
1984 }
1985
1986 /*
1987  * helper function to set a given page up to date if all the
1988  * extents in the tree for that page are up to date
1989  */
1990 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1991 {
1992         u64 start = page_offset(page);
1993         u64 end = start + PAGE_CACHE_SIZE - 1;
1994         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1995                 SetPageUptodate(page);
1996 }
1997
1998 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1999 {
2000         int ret;
2001         int err = 0;
2002         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2003
2004         set_state_private(failure_tree, rec->start, 0);
2005         ret = clear_extent_bits(failure_tree, rec->start,
2006                                 rec->start + rec->len - 1,
2007                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2008         if (ret)
2009                 err = ret;
2010
2011         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
2012                                 rec->start + rec->len - 1,
2013                                 EXTENT_DAMAGED, GFP_NOFS);
2014         if (ret && !err)
2015                 err = ret;
2016
2017         kfree(rec);
2018         return err;
2019 }
2020
2021 /*
2022  * this bypasses the standard btrfs submit functions deliberately, as
2023  * the standard behavior is to write all copies in a raid setup. here we only
2024  * want to write the one bad copy. so we do the mapping for ourselves and issue
2025  * submit_bio directly.
2026  * to avoid any synchronization issues, wait for the data after writing, which
2027  * actually prevents the read that triggered the error from finishing.
2028  * currently, there can be no more than two copies of every data bit. thus,
2029  * exactly one rewrite is required.
2030  */
2031 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2032                       struct page *page, unsigned int pg_offset, int mirror_num)
2033 {
2034         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2035         struct bio *bio;
2036         struct btrfs_device *dev;
2037         u64 map_length = 0;
2038         u64 sector;
2039         struct btrfs_bio *bbio = NULL;
2040         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2041         int ret;
2042
2043         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2044         BUG_ON(!mirror_num);
2045
2046         /* we can't repair anything in raid56 yet */
2047         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2048                 return 0;
2049
2050         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2051         if (!bio)
2052                 return -EIO;
2053         bio->bi_iter.bi_size = 0;
2054         map_length = length;
2055
2056         ret = btrfs_map_block(fs_info, WRITE, logical,
2057                               &map_length, &bbio, mirror_num);
2058         if (ret) {
2059                 bio_put(bio);
2060                 return -EIO;
2061         }
2062         BUG_ON(mirror_num != bbio->mirror_num);
2063         sector = bbio->stripes[mirror_num-1].physical >> 9;
2064         bio->bi_iter.bi_sector = sector;
2065         dev = bbio->stripes[mirror_num-1].dev;
2066         btrfs_put_bbio(bbio);
2067         if (!dev || !dev->bdev || !dev->writeable) {
2068                 bio_put(bio);
2069                 return -EIO;
2070         }
2071         bio->bi_bdev = dev->bdev;
2072         bio_add_page(bio, page, length, pg_offset);
2073
2074         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio)) {
2075                 /* try to remap that extent elsewhere? */
2076                 bio_put(bio);
2077                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2078                 return -EIO;
2079         }
2080
2081         printk_ratelimited_in_rcu(KERN_INFO
2082                                   "BTRFS: read error corrected: ino %llu off %llu (dev %s sector %llu)\n",
2083                                   btrfs_ino(inode), start,
2084                                   rcu_str_deref(dev->name), sector);
2085         bio_put(bio);
2086         return 0;
2087 }
2088
2089 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2090                          int mirror_num)
2091 {
2092         u64 start = eb->start;
2093         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2094         int ret = 0;
2095
2096         if (root->fs_info->sb->s_flags & MS_RDONLY)
2097                 return -EROFS;
2098
2099         for (i = 0; i < num_pages; i++) {
2100                 struct page *p = eb->pages[i];
2101
2102                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2103                                         PAGE_CACHE_SIZE, start, p,
2104                                         start - page_offset(p), mirror_num);
2105                 if (ret)
2106                         break;
2107                 start += PAGE_CACHE_SIZE;
2108         }
2109
2110         return ret;
2111 }
2112
2113 /*
2114  * each time an IO finishes, we do a fast check in the IO failure tree
2115  * to see if we need to process or clean up an io_failure_record
2116  */
2117 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2118                      unsigned int pg_offset)
2119 {
2120         u64 private;
2121         u64 private_failure;
2122         struct io_failure_record *failrec;
2123         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2124         struct extent_state *state;
2125         int num_copies;
2126         int ret;
2127
2128         private = 0;
2129         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2130                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2131         if (!ret)
2132                 return 0;
2133
2134         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2135                                 &private_failure);
2136         if (ret)
2137                 return 0;
2138
2139         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2140         BUG_ON(!failrec->this_mirror);
2141
2142         if (failrec->in_validation) {
2143                 /* there was no real error, just free the record */
2144                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2145                          failrec->start);
2146                 goto out;
2147         }
2148         if (fs_info->sb->s_flags & MS_RDONLY)
2149                 goto out;
2150
2151         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2152         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2153                                             failrec->start,
2154                                             EXTENT_LOCKED);
2155         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2156
2157         if (state && state->start <= failrec->start &&
2158             state->end >= failrec->start + failrec->len - 1) {
2159                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2160                                               failrec->len);
2161                 if (num_copies > 1)  {
2162                         repair_io_failure(inode, start, failrec->len,
2163                                           failrec->logical, page,
2164                                           pg_offset, failrec->failed_mirror);
2165                 }
2166         }
2167
2168 out:
2169         free_io_failure(inode, failrec);
2170
2171         return 0;
2172 }
2173
2174 /*
2175  * Can be called when
2176  * - hold extent lock
2177  * - under ordered extent
2178  * - the inode is freeing
2179  */
2180 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2181 {
2182         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2183         struct io_failure_record *failrec;
2184         struct extent_state *state, *next;
2185
2186         if (RB_EMPTY_ROOT(&failure_tree->state))
2187                 return;
2188
2189         spin_lock(&failure_tree->lock);
2190         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2191         while (state) {
2192                 if (state->start > end)
2193                         break;
2194
2195                 ASSERT(state->end <= end);
2196
2197                 next = next_state(state);
2198
2199                 failrec = (struct io_failure_record *)(unsigned long)state->private;
2200                 free_extent_state(state);
2201                 kfree(failrec);
2202
2203                 state = next;
2204         }
2205         spin_unlock(&failure_tree->lock);
2206 }
2207
2208 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2209                                 struct io_failure_record **failrec_ret)
2210 {
2211         struct io_failure_record *failrec;
2212         u64 private;
2213         struct extent_map *em;
2214         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2215         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2216         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2217         int ret;
2218         u64 logical;
2219
2220         ret = get_state_private(failure_tree, start, &private);
2221         if (ret) {
2222                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2223                 if (!failrec)
2224                         return -ENOMEM;
2225
2226                 failrec->start = start;
2227                 failrec->len = end - start + 1;
2228                 failrec->this_mirror = 0;
2229                 failrec->bio_flags = 0;
2230                 failrec->in_validation = 0;
2231
2232                 read_lock(&em_tree->lock);
2233                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2234                 if (!em) {
2235                         read_unlock(&em_tree->lock);
2236                         kfree(failrec);
2237                         return -EIO;
2238                 }
2239
2240                 if (em->start > start || em->start + em->len <= start) {
2241                         free_extent_map(em);
2242                         em = NULL;
2243                 }
2244                 read_unlock(&em_tree->lock);
2245                 if (!em) {
2246                         kfree(failrec);
2247                         return -EIO;
2248                 }
2249
2250                 logical = start - em->start;
2251                 logical = em->block_start + logical;
2252                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2253                         logical = em->block_start;
2254                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2255                         extent_set_compress_type(&failrec->bio_flags,
2256                                                  em->compress_type);
2257                 }
2258
2259                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2260                          logical, start, failrec->len);
2261
2262                 failrec->logical = logical;
2263                 free_extent_map(em);
2264
2265                 /* set the bits in the private failure tree */
2266                 ret = set_extent_bits(failure_tree, start, end,
2267                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2268                 if (ret >= 0)
2269                         ret = set_state_private(failure_tree, start,
2270                                                 (u64)(unsigned long)failrec);
2271                 /* set the bits in the inode's tree */
2272                 if (ret >= 0)
2273                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2274                                                 GFP_NOFS);
2275                 if (ret < 0) {
2276                         kfree(failrec);
2277                         return ret;
2278                 }
2279         } else {
2280                 failrec = (struct io_failure_record *)(unsigned long)private;
2281                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2282                          failrec->logical, failrec->start, failrec->len,
2283                          failrec->in_validation);
2284                 /*
2285                  * when data can be on disk more than twice, add to failrec here
2286                  * (e.g. with a list for failed_mirror) to make
2287                  * clean_io_failure() clean all those errors at once.
2288                  */
2289         }
2290
2291         *failrec_ret = failrec;
2292
2293         return 0;
2294 }
2295
2296 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2297                            struct io_failure_record *failrec, int failed_mirror)
2298 {
2299         int num_copies;
2300
2301         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2302                                       failrec->logical, failrec->len);
2303         if (num_copies == 1) {
2304                 /*
2305                  * we only have a single copy of the data, so don't bother with
2306                  * all the retry and error correction code that follows. no
2307                  * matter what the error is, it is very likely to persist.
2308                  */
2309                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2310                          num_copies, failrec->this_mirror, failed_mirror);
2311                 return 0;
2312         }
2313
2314         /*
2315          * there are two premises:
2316          *      a) deliver good data to the caller
2317          *      b) correct the bad sectors on disk
2318          */
2319         if (failed_bio->bi_vcnt > 1) {
2320                 /*
2321                  * to fulfill b), we need to know the exact failing sectors, as
2322                  * we don't want to rewrite any more than the failed ones. thus,
2323                  * we need separate read requests for the failed bio
2324                  *
2325                  * if the following BUG_ON triggers, our validation request got
2326                  * merged. we need separate requests for our algorithm to work.
2327                  */
2328                 BUG_ON(failrec->in_validation);
2329                 failrec->in_validation = 1;
2330                 failrec->this_mirror = failed_mirror;
2331         } else {
2332                 /*
2333                  * we're ready to fulfill a) and b) alongside. get a good copy
2334                  * of the failed sector and if we succeed, we have setup
2335                  * everything for repair_io_failure to do the rest for us.
2336                  */
2337                 if (failrec->in_validation) {
2338                         BUG_ON(failrec->this_mirror != failed_mirror);
2339                         failrec->in_validation = 0;
2340                         failrec->this_mirror = 0;
2341                 }
2342                 failrec->failed_mirror = failed_mirror;
2343                 failrec->this_mirror++;
2344                 if (failrec->this_mirror == failed_mirror)
2345                         failrec->this_mirror++;
2346         }
2347
2348         if (failrec->this_mirror > num_copies) {
2349                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2350                          num_copies, failrec->this_mirror, failed_mirror);
2351                 return 0;
2352         }
2353
2354         return 1;
2355 }
2356
2357
2358 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2359                                     struct io_failure_record *failrec,
2360                                     struct page *page, int pg_offset, int icsum,
2361                                     bio_end_io_t *endio_func, void *data)
2362 {
2363         struct bio *bio;
2364         struct btrfs_io_bio *btrfs_failed_bio;
2365         struct btrfs_io_bio *btrfs_bio;
2366
2367         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2368         if (!bio)
2369                 return NULL;
2370
2371         bio->bi_end_io = endio_func;
2372         bio->bi_iter.bi_sector = failrec->logical >> 9;
2373         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2374         bio->bi_iter.bi_size = 0;
2375         bio->bi_private = data;
2376
2377         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2378         if (btrfs_failed_bio->csum) {
2379                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2380                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2381
2382                 btrfs_bio = btrfs_io_bio(bio);
2383                 btrfs_bio->csum = btrfs_bio->csum_inline;
2384                 icsum *= csum_size;
2385                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2386                        csum_size);
2387         }
2388
2389         bio_add_page(bio, page, failrec->len, pg_offset);
2390
2391         return bio;
2392 }
2393
2394 /*
2395  * this is a generic handler for readpage errors (default
2396  * readpage_io_failed_hook). if other copies exist, read those and write back
2397  * good data to the failed position. does not investigate in remapping the
2398  * failed extent elsewhere, hoping the device will be smart enough to do this as
2399  * needed
2400  */
2401
2402 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2403                               struct page *page, u64 start, u64 end,
2404                               int failed_mirror)
2405 {
2406         struct io_failure_record *failrec;
2407         struct inode *inode = page->mapping->host;
2408         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2409         struct bio *bio;
2410         int read_mode;
2411         int ret;
2412
2413         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2414
2415         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2416         if (ret)
2417                 return ret;
2418
2419         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2420         if (!ret) {
2421                 free_io_failure(inode, failrec);
2422                 return -EIO;
2423         }
2424
2425         if (failed_bio->bi_vcnt > 1)
2426                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2427         else
2428                 read_mode = READ_SYNC;
2429
2430         phy_offset >>= inode->i_sb->s_blocksize_bits;
2431         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2432                                       start - page_offset(page),
2433                                       (int)phy_offset, failed_bio->bi_end_io,
2434                                       NULL);
2435         if (!bio) {
2436                 free_io_failure(inode, failrec);
2437                 return -EIO;
2438         }
2439
2440         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2441                  read_mode, failrec->this_mirror, failrec->in_validation);
2442
2443         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2444                                          failrec->this_mirror,
2445                                          failrec->bio_flags, 0);
2446         if (ret) {
2447                 free_io_failure(inode, failrec);
2448                 bio_put(bio);
2449         }
2450
2451         return ret;
2452 }
2453
2454 /* lots and lots of room for performance fixes in the end_bio funcs */
2455
2456 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2457 {
2458         int uptodate = (err == 0);
2459         struct extent_io_tree *tree;
2460         int ret = 0;
2461
2462         tree = &BTRFS_I(page->mapping->host)->io_tree;
2463
2464         if (tree->ops && tree->ops->writepage_end_io_hook) {
2465                 ret = tree->ops->writepage_end_io_hook(page, start,
2466                                                end, NULL, uptodate);
2467                 if (ret)
2468                         uptodate = 0;
2469         }
2470
2471         if (!uptodate) {
2472                 ClearPageUptodate(page);
2473                 SetPageError(page);
2474                 ret = ret < 0 ? ret : -EIO;
2475                 mapping_set_error(page->mapping, ret);
2476         }
2477         return 0;
2478 }
2479
2480 /*
2481  * after a writepage IO is done, we need to:
2482  * clear the uptodate bits on error
2483  * clear the writeback bits in the extent tree for this IO
2484  * end_page_writeback if the page has no more pending IO
2485  *
2486  * Scheduling is not allowed, so the extent state tree is expected
2487  * to have one and only one object corresponding to this IO.
2488  */
2489 static void end_bio_extent_writepage(struct bio *bio, int err)
2490 {
2491         struct bio_vec *bvec;
2492         u64 start;
2493         u64 end;
2494         int i;
2495
2496         bio_for_each_segment_all(bvec, bio, i) {
2497                 struct page *page = bvec->bv_page;
2498
2499                 /* We always issue full-page reads, but if some block
2500                  * in a page fails to read, blk_update_request() will
2501                  * advance bv_offset and adjust bv_len to compensate.
2502                  * Print a warning for nonzero offsets, and an error
2503                  * if they don't add up to a full page.  */
2504                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2505                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2506                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2507                                    "partial page write in btrfs with offset %u and length %u",
2508                                         bvec->bv_offset, bvec->bv_len);
2509                         else
2510                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2511                                    "incomplete page write in btrfs with offset %u and "
2512                                    "length %u",
2513                                         bvec->bv_offset, bvec->bv_len);
2514                 }
2515
2516                 start = page_offset(page);
2517                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2518
2519                 if (end_extent_writepage(page, err, start, end))
2520                         continue;
2521
2522                 end_page_writeback(page);
2523         }
2524
2525         bio_put(bio);
2526 }
2527
2528 static void
2529 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2530                               int uptodate)
2531 {
2532         struct extent_state *cached = NULL;
2533         u64 end = start + len - 1;
2534
2535         if (uptodate && tree->track_uptodate)
2536                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2537         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2538 }
2539
2540 /*
2541  * after a readpage IO is done, we need to:
2542  * clear the uptodate bits on error
2543  * set the uptodate bits if things worked
2544  * set the page up to date if all extents in the tree are uptodate
2545  * clear the lock bit in the extent tree
2546  * unlock the page if there are no other extents locked for it
2547  *
2548  * Scheduling is not allowed, so the extent state tree is expected
2549  * to have one and only one object corresponding to this IO.
2550  */
2551 static void end_bio_extent_readpage(struct bio *bio, int err)
2552 {
2553         struct bio_vec *bvec;
2554         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2555         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2556         struct extent_io_tree *tree;
2557         u64 offset = 0;
2558         u64 start;
2559         u64 end;
2560         u64 len;
2561         u64 extent_start = 0;
2562         u64 extent_len = 0;
2563         int mirror;
2564         int ret;
2565         int i;
2566
2567         if (err)
2568                 uptodate = 0;
2569
2570         bio_for_each_segment_all(bvec, bio, i) {
2571                 struct page *page = bvec->bv_page;
2572                 struct inode *inode = page->mapping->host;
2573
2574                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2575                          "mirror=%u\n", (u64)bio->bi_iter.bi_sector, err,
2576                          io_bio->mirror_num);
2577                 tree = &BTRFS_I(inode)->io_tree;
2578
2579                 /* We always issue full-page reads, but if some block
2580                  * in a page fails to read, blk_update_request() will
2581                  * advance bv_offset and adjust bv_len to compensate.
2582                  * Print a warning for nonzero offsets, and an error
2583                  * if they don't add up to a full page.  */
2584                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2585                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2586                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2587                                    "partial page read in btrfs with offset %u and length %u",
2588                                         bvec->bv_offset, bvec->bv_len);
2589                         else
2590                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2591                                    "incomplete page read in btrfs with offset %u and "
2592                                    "length %u",
2593                                         bvec->bv_offset, bvec->bv_len);
2594                 }
2595
2596                 start = page_offset(page);
2597                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2598                 len = bvec->bv_len;
2599
2600                 mirror = io_bio->mirror_num;
2601                 if (likely(uptodate && tree->ops &&
2602                            tree->ops->readpage_end_io_hook)) {
2603                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2604                                                               page, start, end,
2605                                                               mirror);
2606                         if (ret)
2607                                 uptodate = 0;
2608                         else
2609                                 clean_io_failure(inode, start, page, 0);
2610                 }
2611
2612                 if (likely(uptodate))
2613                         goto readpage_ok;
2614
2615                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2616                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2617                         if (!ret && !err &&
2618                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2619                                 uptodate = 1;
2620                 } else {
2621                         /*
2622                          * The generic bio_readpage_error handles errors the
2623                          * following way: If possible, new read requests are
2624                          * created and submitted and will end up in
2625                          * end_bio_extent_readpage as well (if we're lucky, not
2626                          * in the !uptodate case). In that case it returns 0 and
2627                          * we just go on with the next page in our bio. If it
2628                          * can't handle the error it will return -EIO and we
2629                          * remain responsible for that page.
2630                          */
2631                         ret = bio_readpage_error(bio, offset, page, start, end,
2632                                                  mirror);
2633                         if (ret == 0) {
2634                                 uptodate =
2635                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2636                                 if (err)
2637                                         uptodate = 0;
2638                                 offset += len;
2639                                 continue;
2640                         }
2641                 }
2642 readpage_ok:
2643                 if (likely(uptodate)) {
2644                         loff_t i_size = i_size_read(inode);
2645                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2646                         unsigned off;
2647
2648                         /* Zero out the end if this page straddles i_size */
2649                         off = i_size & (PAGE_CACHE_SIZE-1);
2650                         if (page->index == end_index && off)
2651                                 zero_user_segment(page, off, PAGE_CACHE_SIZE);
2652                         SetPageUptodate(page);
2653                 } else {
2654                         ClearPageUptodate(page);
2655                         SetPageError(page);
2656                 }
2657                 unlock_page(page);
2658                 offset += len;
2659
2660                 if (unlikely(!uptodate)) {
2661                         if (extent_len) {
2662                                 endio_readpage_release_extent(tree,
2663                                                               extent_start,
2664                                                               extent_len, 1);
2665                                 extent_start = 0;
2666                                 extent_len = 0;
2667                         }
2668                         endio_readpage_release_extent(tree, start,
2669                                                       end - start + 1, 0);
2670                 } else if (!extent_len) {
2671                         extent_start = start;
2672                         extent_len = end + 1 - start;
2673                 } else if (extent_start + extent_len == start) {
2674                         extent_len += end + 1 - start;
2675                 } else {
2676                         endio_readpage_release_extent(tree, extent_start,
2677                                                       extent_len, uptodate);
2678                         extent_start = start;
2679                         extent_len = end + 1 - start;
2680                 }
2681         }
2682
2683         if (extent_len)
2684                 endio_readpage_release_extent(tree, extent_start, extent_len,
2685                                               uptodate);
2686         if (io_bio->end_io)
2687                 io_bio->end_io(io_bio, err);
2688         bio_put(bio);
2689 }
2690
2691 /*
2692  * this allocates from the btrfs_bioset.  We're returning a bio right now
2693  * but you can call btrfs_io_bio for the appropriate container_of magic
2694  */
2695 struct bio *
2696 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2697                 gfp_t gfp_flags)
2698 {
2699         struct btrfs_io_bio *btrfs_bio;
2700         struct bio *bio;
2701
2702         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2703
2704         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2705                 while (!bio && (nr_vecs /= 2)) {
2706                         bio = bio_alloc_bioset(gfp_flags,
2707                                                nr_vecs, btrfs_bioset);
2708                 }
2709         }
2710
2711         if (bio) {
2712                 bio->bi_bdev = bdev;
2713                 bio->bi_iter.bi_sector = first_sector;
2714                 btrfs_bio = btrfs_io_bio(bio);
2715                 btrfs_bio->csum = NULL;
2716                 btrfs_bio->csum_allocated = NULL;
2717                 btrfs_bio->end_io = NULL;
2718         }
2719         return bio;
2720 }
2721
2722 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2723 {
2724         struct btrfs_io_bio *btrfs_bio;
2725         struct bio *new;
2726
2727         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2728         if (new) {
2729                 btrfs_bio = btrfs_io_bio(new);
2730                 btrfs_bio->csum = NULL;
2731                 btrfs_bio->csum_allocated = NULL;
2732                 btrfs_bio->end_io = NULL;
2733                 /* FIXME, put this into bio_clone_bioset */
2734                 if (bio->bi_css)
2735                         bio_associate_blkcg(new, bio->bi_css);
2736         }
2737         return new;
2738 }
2739
2740 /* this also allocates from the btrfs_bioset */
2741 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2742 {
2743         struct btrfs_io_bio *btrfs_bio;
2744         struct bio *bio;
2745
2746         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2747         if (bio) {
2748                 btrfs_bio = btrfs_io_bio(bio);
2749                 btrfs_bio->csum = NULL;
2750                 btrfs_bio->csum_allocated = NULL;
2751                 btrfs_bio->end_io = NULL;
2752         }
2753         return bio;
2754 }
2755
2756
2757 static int __must_check submit_one_bio(int rw, struct bio *bio,
2758                                        int mirror_num, unsigned long bio_flags)
2759 {
2760         int ret = 0;
2761         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2762         struct page *page = bvec->bv_page;
2763         struct extent_io_tree *tree = bio->bi_private;
2764         u64 start;
2765
2766         start = page_offset(page) + bvec->bv_offset;
2767
2768         bio->bi_private = NULL;
2769
2770         bio_get(bio);
2771
2772         if (tree->ops && tree->ops->submit_bio_hook)
2773                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2774                                            mirror_num, bio_flags, start);
2775         else
2776                 btrfsic_submit_bio(rw, bio);
2777
2778         bio_put(bio);
2779         return ret;
2780 }
2781
2782 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2783                      unsigned long offset, size_t size, struct bio *bio,
2784                      unsigned long bio_flags)
2785 {
2786         int ret = 0;
2787         if (tree->ops && tree->ops->merge_bio_hook)
2788                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2789                                                 bio_flags);
2790         BUG_ON(ret < 0);
2791         return ret;
2792
2793 }
2794
2795 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2796                               struct writeback_control *wbc,
2797                               struct page *page, sector_t sector,
2798                               size_t size, unsigned long offset,
2799                               struct block_device *bdev,
2800                               struct bio **bio_ret,
2801                               unsigned long max_pages,
2802                               bio_end_io_t end_io_func,
2803                               int mirror_num,
2804                               unsigned long prev_bio_flags,
2805                               unsigned long bio_flags)
2806 {
2807         int ret = 0;
2808         struct bio *bio;
2809         int nr;
2810         int contig = 0;
2811         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2812         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2813         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2814
2815         if (bio_ret && *bio_ret) {
2816                 bio = *bio_ret;
2817                 if (old_compressed)
2818                         contig = bio->bi_iter.bi_sector == sector;
2819                 else
2820                         contig = bio_end_sector(bio) == sector;
2821
2822                 if (prev_bio_flags != bio_flags || !contig ||
2823                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2824                     bio_add_page(bio, page, page_size, offset) < page_size) {
2825                         ret = submit_one_bio(rw, bio, mirror_num,
2826                                              prev_bio_flags);
2827                         if (ret < 0) {
2828                                 *bio_ret = NULL;
2829                                 return ret;
2830                         }
2831                         bio = NULL;
2832                 } else {
2833                         if (wbc)
2834                                 wbc_account_io(wbc, page, page_size);
2835                         return 0;
2836                 }
2837         }
2838         if (this_compressed)
2839                 nr = BIO_MAX_PAGES;
2840         else
2841                 nr = bio_get_nr_vecs(bdev);
2842
2843         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2844         if (!bio)
2845                 return -ENOMEM;
2846
2847         bio_add_page(bio, page, page_size, offset);
2848         bio->bi_end_io = end_io_func;
2849         bio->bi_private = tree;
2850         if (wbc) {
2851                 wbc_init_bio(wbc, bio);
2852                 wbc_account_io(wbc, page, page_size);
2853         }
2854
2855         if (bio_ret)
2856                 *bio_ret = bio;
2857         else
2858                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2859
2860         return ret;
2861 }
2862
2863 static void attach_extent_buffer_page(struct extent_buffer *eb,
2864                                       struct page *page)
2865 {
2866         if (!PagePrivate(page)) {
2867                 SetPagePrivate(page);
2868                 page_cache_get(page);
2869                 set_page_private(page, (unsigned long)eb);
2870         } else {
2871                 WARN_ON(page->private != (unsigned long)eb);
2872         }
2873 }
2874
2875 void set_page_extent_mapped(struct page *page)
2876 {
2877         if (!PagePrivate(page)) {
2878                 SetPagePrivate(page);
2879                 page_cache_get(page);
2880                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2881         }
2882 }
2883
2884 static struct extent_map *
2885 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2886                  u64 start, u64 len, get_extent_t *get_extent,
2887                  struct extent_map **em_cached)
2888 {
2889         struct extent_map *em;
2890
2891         if (em_cached && *em_cached) {
2892                 em = *em_cached;
2893                 if (extent_map_in_tree(em) && start >= em->start &&
2894                     start < extent_map_end(em)) {
2895                         atomic_inc(&em->refs);
2896                         return em;
2897                 }
2898
2899                 free_extent_map(em);
2900                 *em_cached = NULL;
2901         }
2902
2903         em = get_extent(inode, page, pg_offset, start, len, 0);
2904         if (em_cached && !IS_ERR_OR_NULL(em)) {
2905                 BUG_ON(*em_cached);
2906                 atomic_inc(&em->refs);
2907                 *em_cached = em;
2908         }
2909         return em;
2910 }
2911 /*
2912  * basic readpage implementation.  Locked extent state structs are inserted
2913  * into the tree that are removed when the IO is done (by the end_io
2914  * handlers)
2915  * XXX JDM: This needs looking at to ensure proper page locking
2916  */
2917 static int __do_readpage(struct extent_io_tree *tree,
2918                          struct page *page,
2919                          get_extent_t *get_extent,
2920                          struct extent_map **em_cached,
2921                          struct bio **bio, int mirror_num,
2922                          unsigned long *bio_flags, int rw)
2923 {
2924         struct inode *inode = page->mapping->host;
2925         u64 start = page_offset(page);
2926         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2927         u64 end;
2928         u64 cur = start;
2929         u64 extent_offset;
2930         u64 last_byte = i_size_read(inode);
2931         u64 block_start;
2932         u64 cur_end;
2933         sector_t sector;
2934         struct extent_map *em;
2935         struct block_device *bdev;
2936         int ret;
2937         int nr = 0;
2938         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2939         size_t pg_offset = 0;
2940         size_t iosize;
2941         size_t disk_io_size;
2942         size_t blocksize = inode->i_sb->s_blocksize;
2943         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2944
2945         set_page_extent_mapped(page);
2946
2947         end = page_end;
2948         if (!PageUptodate(page)) {
2949                 if (cleancache_get_page(page) == 0) {
2950                         BUG_ON(blocksize != PAGE_SIZE);
2951                         unlock_extent(tree, start, end);
2952                         goto out;
2953                 }
2954         }
2955
2956         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2957                 char *userpage;
2958                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2959
2960                 if (zero_offset) {
2961                         iosize = PAGE_CACHE_SIZE - zero_offset;
2962                         userpage = kmap_atomic(page);
2963                         memset(userpage + zero_offset, 0, iosize);
2964                         flush_dcache_page(page);
2965                         kunmap_atomic(userpage);
2966                 }
2967         }
2968         while (cur <= end) {
2969                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2970
2971                 if (cur >= last_byte) {
2972                         char *userpage;
2973                         struct extent_state *cached = NULL;
2974
2975                         iosize = PAGE_CACHE_SIZE - pg_offset;
2976                         userpage = kmap_atomic(page);
2977                         memset(userpage + pg_offset, 0, iosize);
2978                         flush_dcache_page(page);
2979                         kunmap_atomic(userpage);
2980                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2981                                             &cached, GFP_NOFS);
2982                         if (!parent_locked)
2983                                 unlock_extent_cached(tree, cur,
2984                                                      cur + iosize - 1,
2985                                                      &cached, GFP_NOFS);
2986                         break;
2987                 }
2988                 em = __get_extent_map(inode, page, pg_offset, cur,
2989                                       end - cur + 1, get_extent, em_cached);
2990                 if (IS_ERR_OR_NULL(em)) {
2991                         SetPageError(page);
2992                         if (!parent_locked)
2993                                 unlock_extent(tree, cur, end);
2994                         break;
2995                 }
2996                 extent_offset = cur - em->start;
2997                 BUG_ON(extent_map_end(em) <= cur);
2998                 BUG_ON(end < cur);
2999
3000                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
3001                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
3002                         extent_set_compress_type(&this_bio_flag,
3003                                                  em->compress_type);
3004                 }
3005
3006                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
3007                 cur_end = min(extent_map_end(em) - 1, end);
3008                 iosize = ALIGN(iosize, blocksize);
3009                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
3010                         disk_io_size = em->block_len;
3011                         sector = em->block_start >> 9;
3012                 } else {
3013                         sector = (em->block_start + extent_offset) >> 9;
3014                         disk_io_size = iosize;
3015                 }
3016                 bdev = em->bdev;
3017                 block_start = em->block_start;
3018                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3019                         block_start = EXTENT_MAP_HOLE;
3020                 free_extent_map(em);
3021                 em = NULL;
3022
3023                 /* we've found a hole, just zero and go on */
3024                 if (block_start == EXTENT_MAP_HOLE) {
3025                         char *userpage;
3026                         struct extent_state *cached = NULL;
3027
3028                         userpage = kmap_atomic(page);
3029                         memset(userpage + pg_offset, 0, iosize);
3030                         flush_dcache_page(page);
3031                         kunmap_atomic(userpage);
3032
3033                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3034                                             &cached, GFP_NOFS);
3035                         unlock_extent_cached(tree, cur, cur + iosize - 1,
3036                                              &cached, GFP_NOFS);
3037                         cur = cur + iosize;
3038                         pg_offset += iosize;
3039                         continue;
3040                 }
3041                 /* the get_extent function already copied into the page */
3042                 if (test_range_bit(tree, cur, cur_end,
3043                                    EXTENT_UPTODATE, 1, NULL)) {
3044                         check_page_uptodate(tree, page);
3045                         if (!parent_locked)
3046                                 unlock_extent(tree, cur, cur + iosize - 1);
3047                         cur = cur + iosize;
3048                         pg_offset += iosize;
3049                         continue;
3050                 }
3051                 /* we have an inline extent but it didn't get marked up
3052                  * to date.  Error out
3053                  */
3054                 if (block_start == EXTENT_MAP_INLINE) {
3055                         SetPageError(page);
3056                         if (!parent_locked)
3057                                 unlock_extent(tree, cur, cur + iosize - 1);
3058                         cur = cur + iosize;
3059                         pg_offset += iosize;
3060                         continue;
3061                 }
3062
3063                 pnr -= page->index;
3064                 ret = submit_extent_page(rw, tree, NULL, page,
3065                                          sector, disk_io_size, pg_offset,
3066                                          bdev, bio, pnr,
3067                                          end_bio_extent_readpage, mirror_num,
3068                                          *bio_flags,
3069                                          this_bio_flag);
3070                 if (!ret) {
3071                         nr++;
3072                         *bio_flags = this_bio_flag;
3073                 } else {
3074                         SetPageError(page);
3075                         if (!parent_locked)
3076                                 unlock_extent(tree, cur, cur + iosize - 1);
3077                 }
3078                 cur = cur + iosize;
3079                 pg_offset += iosize;
3080         }
3081 out:
3082         if (!nr) {
3083                 if (!PageError(page))
3084                         SetPageUptodate(page);
3085                 unlock_page(page);
3086         }
3087         return 0;
3088 }
3089
3090 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3091                                              struct page *pages[], int nr_pages,
3092                                              u64 start, u64 end,
3093                                              get_extent_t *get_extent,
3094                                              struct extent_map **em_cached,
3095                                              struct bio **bio, int mirror_num,
3096                                              unsigned long *bio_flags, int rw)
3097 {
3098         struct inode *inode;
3099         struct btrfs_ordered_extent *ordered;
3100         int index;
3101
3102         inode = pages[0]->mapping->host;
3103         while (1) {
3104                 lock_extent(tree, start, end);
3105                 ordered = btrfs_lookup_ordered_range(inode, start,
3106                                                      end - start + 1);
3107                 if (!ordered)
3108                         break;
3109                 unlock_extent(tree, start, end);
3110                 btrfs_start_ordered_extent(inode, ordered, 1);
3111                 btrfs_put_ordered_extent(ordered);
3112         }
3113
3114         for (index = 0; index < nr_pages; index++) {
3115                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3116                               mirror_num, bio_flags, rw);
3117                 page_cache_release(pages[index]);
3118         }
3119 }
3120
3121 static void __extent_readpages(struct extent_io_tree *tree,
3122                                struct page *pages[],
3123                                int nr_pages, get_extent_t *get_extent,
3124                                struct extent_map **em_cached,
3125                                struct bio **bio, int mirror_num,
3126                                unsigned long *bio_flags, int rw)
3127 {
3128         u64 start = 0;
3129         u64 end = 0;
3130         u64 page_start;
3131         int index;
3132         int first_index = 0;
3133
3134         for (index = 0; index < nr_pages; index++) {
3135                 page_start = page_offset(pages[index]);
3136                 if (!end) {
3137                         start = page_start;
3138                         end = start + PAGE_CACHE_SIZE - 1;
3139                         first_index = index;
3140                 } else if (end + 1 == page_start) {
3141                         end += PAGE_CACHE_SIZE;
3142                 } else {
3143                         __do_contiguous_readpages(tree, &pages[first_index],
3144                                                   index - first_index, start,
3145                                                   end, get_extent, em_cached,
3146                                                   bio, mirror_num, bio_flags,
3147                                                   rw);
3148                         start = page_start;
3149                         end = start + PAGE_CACHE_SIZE - 1;
3150                         first_index = index;
3151                 }
3152         }
3153
3154         if (end)
3155                 __do_contiguous_readpages(tree, &pages[first_index],
3156                                           index - first_index, start,
3157                                           end, get_extent, em_cached, bio,
3158                                           mirror_num, bio_flags, rw);
3159 }
3160
3161 static int __extent_read_full_page(struct extent_io_tree *tree,
3162                                    struct page *page,
3163                                    get_extent_t *get_extent,
3164                                    struct bio **bio, int mirror_num,
3165                                    unsigned long *bio_flags, int rw)
3166 {
3167         struct inode *inode = page->mapping->host;
3168         struct btrfs_ordered_extent *ordered;
3169         u64 start = page_offset(page);
3170         u64 end = start + PAGE_CACHE_SIZE - 1;
3171         int ret;
3172
3173         while (1) {
3174                 lock_extent(tree, start, end);
3175                 ordered = btrfs_lookup_ordered_extent(inode, start);
3176                 if (!ordered)
3177                         break;
3178                 unlock_extent(tree, start, end);
3179                 btrfs_start_ordered_extent(inode, ordered, 1);
3180                 btrfs_put_ordered_extent(ordered);
3181         }
3182
3183         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3184                             bio_flags, rw);
3185         return ret;
3186 }
3187
3188 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3189                             get_extent_t *get_extent, int mirror_num)
3190 {
3191         struct bio *bio = NULL;
3192         unsigned long bio_flags = 0;
3193         int ret;
3194
3195         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3196                                       &bio_flags, READ);
3197         if (bio)
3198                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3199         return ret;
3200 }
3201
3202 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3203                                  get_extent_t *get_extent, int mirror_num)
3204 {
3205         struct bio *bio = NULL;
3206         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3207         int ret;
3208
3209         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3210                                       &bio_flags, READ);
3211         if (bio)
3212                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3213         return ret;
3214 }
3215
3216 static noinline void update_nr_written(struct page *page,
3217                                       struct writeback_control *wbc,
3218                                       unsigned long nr_written)
3219 {
3220         wbc->nr_to_write -= nr_written;
3221         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3222             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3223                 page->mapping->writeback_index = page->index + nr_written;
3224 }
3225
3226 /*
3227  * helper for __extent_writepage, doing all of the delayed allocation setup.
3228  *
3229  * This returns 1 if our fill_delalloc function did all the work required
3230  * to write the page (copy into inline extent).  In this case the IO has
3231  * been started and the page is already unlocked.
3232  *
3233  * This returns 0 if all went well (page still locked)
3234  * This returns < 0 if there were errors (page still locked)
3235  */
3236 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3237                               struct page *page, struct writeback_control *wbc,
3238                               struct extent_page_data *epd,
3239                               u64 delalloc_start,
3240                               unsigned long *nr_written)
3241 {
3242         struct extent_io_tree *tree = epd->tree;
3243         u64 page_end = delalloc_start + PAGE_CACHE_SIZE - 1;
3244         u64 nr_delalloc;
3245         u64 delalloc_to_write = 0;
3246         u64 delalloc_end = 0;
3247         int ret;
3248         int page_started = 0;
3249
3250         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3251                 return 0;
3252
3253         while (delalloc_end < page_end) {
3254                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3255                                                page,
3256                                                &delalloc_start,
3257                                                &delalloc_end,
3258                                                BTRFS_MAX_EXTENT_SIZE);
3259                 if (nr_delalloc == 0) {
3260                         delalloc_start = delalloc_end + 1;
3261                         continue;
3262                 }
3263                 ret = tree->ops->fill_delalloc(inode, page,
3264                                                delalloc_start,
3265                                                delalloc_end,
3266                                                &page_started,
3267                                                nr_written);
3268                 /* File system has been set read-only */
3269                 if (ret) {
3270                         SetPageError(page);
3271                         /* fill_delalloc should be return < 0 for error
3272                          * but just in case, we use > 0 here meaning the
3273                          * IO is started, so we don't want to return > 0
3274                          * unless things are going well.
3275                          */
3276                         ret = ret < 0 ? ret : -EIO;
3277                         goto done;
3278                 }
3279                 /*
3280                  * delalloc_end is already one less than the total
3281                  * length, so we don't subtract one from
3282                  * PAGE_CACHE_SIZE
3283                  */
3284                 delalloc_to_write += (delalloc_end - delalloc_start +
3285                                       PAGE_CACHE_SIZE) >>
3286                                       PAGE_CACHE_SHIFT;
3287                 delalloc_start = delalloc_end + 1;
3288         }
3289         if (wbc->nr_to_write < delalloc_to_write) {
3290                 int thresh = 8192;
3291
3292                 if (delalloc_to_write < thresh * 2)
3293                         thresh = delalloc_to_write;
3294                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3295                                          thresh);
3296         }
3297
3298         /* did the fill delalloc function already unlock and start
3299          * the IO?
3300          */
3301         if (page_started) {
3302                 /*
3303                  * we've unlocked the page, so we can't update
3304                  * the mapping's writeback index, just update
3305                  * nr_to_write.
3306                  */
3307                 wbc->nr_to_write -= *nr_written;
3308                 return 1;
3309         }
3310
3311         ret = 0;
3312
3313 done:
3314         return ret;
3315 }
3316
3317 /*
3318  * helper for __extent_writepage.  This calls the writepage start hooks,
3319  * and does the loop to map the page into extents and bios.
3320  *
3321  * We return 1 if the IO is started and the page is unlocked,
3322  * 0 if all went well (page still locked)
3323  * < 0 if there were errors (page still locked)
3324  */
3325 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3326                                  struct page *page,
3327                                  struct writeback_control *wbc,
3328                                  struct extent_page_data *epd,
3329                                  loff_t i_size,
3330                                  unsigned long nr_written,
3331                                  int write_flags, int *nr_ret)
3332 {
3333         struct extent_io_tree *tree = epd->tree;
3334         u64 start = page_offset(page);
3335         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3336         u64 end;
3337         u64 cur = start;
3338         u64 extent_offset;
3339         u64 block_start;
3340         u64 iosize;
3341         sector_t sector;
3342         struct extent_state *cached_state = NULL;
3343         struct extent_map *em;
3344         struct block_device *bdev;
3345         size_t pg_offset = 0;
3346         size_t blocksize;
3347         int ret = 0;
3348         int nr = 0;
3349         bool compressed;
3350
3351         if (tree->ops && tree->ops->writepage_start_hook) {
3352                 ret = tree->ops->writepage_start_hook(page, start,
3353                                                       page_end);
3354                 if (ret) {
3355                         /* Fixup worker will requeue */
3356                         if (ret == -EBUSY)
3357                                 wbc->pages_skipped++;
3358                         else
3359                                 redirty_page_for_writepage(wbc, page);
3360
3361                         update_nr_written(page, wbc, nr_written);
3362                         unlock_page(page);
3363                         ret = 1;
3364                         goto done_unlocked;
3365                 }
3366         }
3367
3368         /*
3369          * we don't want to touch the inode after unlocking the page,
3370          * so we update the mapping writeback index now
3371          */
3372         update_nr_written(page, wbc, nr_written + 1);
3373
3374         end = page_end;
3375         if (i_size <= start) {
3376                 if (tree->ops && tree->ops->writepage_end_io_hook)
3377                         tree->ops->writepage_end_io_hook(page, start,
3378                                                          page_end, NULL, 1);
3379                 goto done;
3380         }
3381
3382         blocksize = inode->i_sb->s_blocksize;
3383
3384         while (cur <= end) {
3385                 u64 em_end;
3386                 if (cur >= i_size) {
3387                         if (tree->ops && tree->ops->writepage_end_io_hook)
3388                                 tree->ops->writepage_end_io_hook(page, cur,
3389                                                          page_end, NULL, 1);
3390                         break;
3391                 }
3392                 em = epd->get_extent(inode, page, pg_offset, cur,
3393                                      end - cur + 1, 1);
3394                 if (IS_ERR_OR_NULL(em)) {
3395                         SetPageError(page);
3396                         ret = PTR_ERR_OR_ZERO(em);
3397                         break;
3398                 }
3399
3400                 extent_offset = cur - em->start;
3401                 em_end = extent_map_end(em);
3402                 BUG_ON(em_end <= cur);
3403                 BUG_ON(end < cur);
3404                 iosize = min(em_end - cur, end - cur + 1);
3405                 iosize = ALIGN(iosize, blocksize);
3406                 sector = (em->block_start + extent_offset) >> 9;
3407                 bdev = em->bdev;
3408                 block_start = em->block_start;
3409                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3410                 free_extent_map(em);
3411                 em = NULL;
3412
3413                 /*
3414                  * compressed and inline extents are written through other
3415                  * paths in the FS
3416                  */
3417                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3418                     block_start == EXTENT_MAP_INLINE) {
3419                         /*
3420                          * end_io notification does not happen here for
3421                          * compressed extents
3422                          */
3423                         if (!compressed && tree->ops &&
3424                             tree->ops->writepage_end_io_hook)
3425                                 tree->ops->writepage_end_io_hook(page, cur,
3426                                                          cur + iosize - 1,
3427                                                          NULL, 1);
3428                         else if (compressed) {
3429                                 /* we don't want to end_page_writeback on
3430                                  * a compressed extent.  this happens
3431                                  * elsewhere
3432                                  */
3433                                 nr++;
3434                         }
3435
3436                         cur += iosize;
3437                         pg_offset += iosize;
3438                         continue;
3439                 }
3440
3441                 if (tree->ops && tree->ops->writepage_io_hook) {
3442                         ret = tree->ops->writepage_io_hook(page, cur,
3443                                                 cur + iosize - 1);
3444                 } else {
3445                         ret = 0;
3446                 }
3447                 if (ret) {
3448                         SetPageError(page);
3449                 } else {
3450                         unsigned long max_nr = (i_size >> PAGE_CACHE_SHIFT) + 1;
3451
3452                         set_range_writeback(tree, cur, cur + iosize - 1);
3453                         if (!PageWriteback(page)) {
3454                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
3455                                            "page %lu not writeback, cur %llu end %llu",
3456                                        page->index, cur, end);
3457                         }
3458
3459                         ret = submit_extent_page(write_flags, tree, wbc, page,
3460                                                  sector, iosize, pg_offset,
3461                                                  bdev, &epd->bio, max_nr,
3462                                                  end_bio_extent_writepage,
3463                                                  0, 0, 0);
3464                         if (ret)
3465                                 SetPageError(page);
3466                 }
3467                 cur = cur + iosize;
3468                 pg_offset += iosize;
3469                 nr++;
3470         }
3471 done:
3472         *nr_ret = nr;
3473
3474 done_unlocked:
3475
3476         /* drop our reference on any cached states */
3477         free_extent_state(cached_state);
3478         return ret;
3479 }
3480
3481 /*
3482  * the writepage semantics are similar to regular writepage.  extent
3483  * records are inserted to lock ranges in the tree, and as dirty areas
3484  * are found, they are marked writeback.  Then the lock bits are removed
3485  * and the end_io handler clears the writeback ranges
3486  */
3487 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3488                               void *data)
3489 {
3490         struct inode *inode = page->mapping->host;
3491         struct extent_page_data *epd = data;
3492         u64 start = page_offset(page);
3493         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3494         int ret;
3495         int nr = 0;
3496         size_t pg_offset = 0;
3497         loff_t i_size = i_size_read(inode);
3498         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3499         int write_flags;
3500         unsigned long nr_written = 0;
3501
3502         if (wbc->sync_mode == WB_SYNC_ALL)
3503                 write_flags = WRITE_SYNC;
3504         else
3505                 write_flags = WRITE;
3506
3507         trace___extent_writepage(page, inode, wbc);
3508
3509         WARN_ON(!PageLocked(page));
3510
3511         ClearPageError(page);
3512
3513         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3514         if (page->index > end_index ||
3515            (page->index == end_index && !pg_offset)) {
3516                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3517                 unlock_page(page);
3518                 return 0;
3519         }
3520
3521         if (page->index == end_index) {
3522                 char *userpage;
3523
3524                 userpage = kmap_atomic(page);
3525                 memset(userpage + pg_offset, 0,
3526                        PAGE_CACHE_SIZE - pg_offset);
3527                 kunmap_atomic(userpage);
3528                 flush_dcache_page(page);
3529         }
3530
3531         pg_offset = 0;
3532
3533         set_page_extent_mapped(page);
3534
3535         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3536         if (ret == 1)
3537                 goto done_unlocked;
3538         if (ret)
3539                 goto done;
3540
3541         ret = __extent_writepage_io(inode, page, wbc, epd,
3542                                     i_size, nr_written, write_flags, &nr);
3543         if (ret == 1)
3544                 goto done_unlocked;
3545
3546 done:
3547         if (nr == 0) {
3548                 /* make sure the mapping tag for page dirty gets cleared */
3549                 set_page_writeback(page);
3550                 end_page_writeback(page);
3551         }
3552         if (PageError(page)) {
3553                 ret = ret < 0 ? ret : -EIO;
3554                 end_extent_writepage(page, ret, start, page_end);
3555         }
3556         unlock_page(page);
3557         return ret;
3558
3559 done_unlocked:
3560         return 0;
3561 }
3562
3563 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3564 {
3565         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3566                        TASK_UNINTERRUPTIBLE);
3567 }
3568
3569 static noinline_for_stack int
3570 lock_extent_buffer_for_io(struct extent_buffer *eb,
3571                           struct btrfs_fs_info *fs_info,
3572                           struct extent_page_data *epd)
3573 {
3574         unsigned long i, num_pages;
3575         int flush = 0;
3576         int ret = 0;
3577
3578         if (!btrfs_try_tree_write_lock(eb)) {
3579                 flush = 1;
3580                 flush_write_bio(epd);
3581                 btrfs_tree_lock(eb);
3582         }
3583
3584         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3585                 btrfs_tree_unlock(eb);
3586                 if (!epd->sync_io)
3587                         return 0;
3588                 if (!flush) {
3589                         flush_write_bio(epd);
3590                         flush = 1;
3591                 }
3592                 while (1) {
3593                         wait_on_extent_buffer_writeback(eb);
3594                         btrfs_tree_lock(eb);
3595                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3596                                 break;
3597                         btrfs_tree_unlock(eb);
3598                 }
3599         }
3600
3601         /*
3602          * We need to do this to prevent races in people who check if the eb is
3603          * under IO since we can end up having no IO bits set for a short period
3604          * of time.
3605          */
3606         spin_lock(&eb->refs_lock);
3607         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3608                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3609                 spin_unlock(&eb->refs_lock);
3610                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3611                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3612                                      -eb->len,
3613                                      fs_info->dirty_metadata_batch);
3614                 ret = 1;
3615         } else {
3616                 spin_unlock(&eb->refs_lock);
3617         }
3618
3619         btrfs_tree_unlock(eb);
3620
3621         if (!ret)
3622                 return ret;
3623
3624         num_pages = num_extent_pages(eb->start, eb->len);
3625         for (i = 0; i < num_pages; i++) {
3626                 struct page *p = eb->pages[i];
3627
3628                 if (!trylock_page(p)) {
3629                         if (!flush) {
3630                                 flush_write_bio(epd);
3631                                 flush = 1;
3632                         }
3633                         lock_page(p);
3634                 }
3635         }
3636
3637         return ret;
3638 }
3639
3640 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3641 {
3642         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3643         smp_mb__after_atomic();
3644         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3645 }
3646
3647 static void set_btree_ioerr(struct page *page)
3648 {
3649         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3650         struct btrfs_inode *btree_ino = BTRFS_I(eb->fs_info->btree_inode);
3651
3652         SetPageError(page);
3653         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3654                 return;
3655
3656         /*
3657          * If writeback for a btree extent that doesn't belong to a log tree
3658          * failed, increment the counter transaction->eb_write_errors.
3659          * We do this because while the transaction is running and before it's
3660          * committing (when we call filemap_fdata[write|wait]_range against
3661          * the btree inode), we might have
3662          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3663          * returns an error or an error happens during writeback, when we're
3664          * committing the transaction we wouldn't know about it, since the pages
3665          * can be no longer dirty nor marked anymore for writeback (if a
3666          * subsequent modification to the extent buffer didn't happen before the
3667          * transaction commit), which makes filemap_fdata[write|wait]_range not
3668          * able to find the pages tagged with SetPageError at transaction
3669          * commit time. So if this happens we must abort the transaction,
3670          * otherwise we commit a super block with btree roots that point to
3671          * btree nodes/leafs whose content on disk is invalid - either garbage
3672          * or the content of some node/leaf from a past generation that got
3673          * cowed or deleted and is no longer valid.
3674          *
3675          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3676          * not be enough - we need to distinguish between log tree extents vs
3677          * non-log tree extents, and the next filemap_fdatawait_range() call
3678          * will catch and clear such errors in the mapping - and that call might
3679          * be from a log sync and not from a transaction commit. Also, checking
3680          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3681          * not done and would not be reliable - the eb might have been released
3682          * from memory and reading it back again means that flag would not be
3683          * set (since it's a runtime flag, not persisted on disk).
3684          *
3685          * Using the flags below in the btree inode also makes us achieve the
3686          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3687          * writeback for all dirty pages and before filemap_fdatawait_range()
3688          * is called, the writeback for all dirty pages had already finished
3689          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3690          * filemap_fdatawait_range() would return success, as it could not know
3691          * that writeback errors happened (the pages were no longer tagged for
3692          * writeback).
3693          */
3694         switch (eb->log_index) {
3695         case -1:
3696                 set_bit(BTRFS_INODE_BTREE_ERR, &btree_ino->runtime_flags);
3697                 break;
3698         case 0:
3699                 set_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
3700                 break;
3701         case 1:
3702                 set_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
3703                 break;
3704         default:
3705                 BUG(); /* unexpected, logic error */
3706         }
3707 }
3708
3709 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3710 {
3711         struct bio_vec *bvec;
3712         struct extent_buffer *eb;
3713         int i, done;
3714
3715         bio_for_each_segment_all(bvec, bio, i) {
3716                 struct page *page = bvec->bv_page;
3717
3718                 eb = (struct extent_buffer *)page->private;
3719                 BUG_ON(!eb);
3720                 done = atomic_dec_and_test(&eb->io_pages);
3721
3722                 if (err || test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3723                         ClearPageUptodate(page);
3724                         set_btree_ioerr(page);
3725                 }
3726
3727                 end_page_writeback(page);
3728
3729                 if (!done)
3730                         continue;
3731
3732                 end_extent_buffer_writeback(eb);
3733         }
3734
3735         bio_put(bio);
3736 }
3737
3738 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3739                         struct btrfs_fs_info *fs_info,
3740                         struct writeback_control *wbc,
3741                         struct extent_page_data *epd)
3742 {
3743         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3744         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3745         u64 offset = eb->start;
3746         unsigned long i, num_pages;
3747         unsigned long bio_flags = 0;
3748         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3749         int ret = 0;
3750
3751         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3752         num_pages = num_extent_pages(eb->start, eb->len);
3753         atomic_set(&eb->io_pages, num_pages);
3754         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3755                 bio_flags = EXTENT_BIO_TREE_LOG;
3756
3757         for (i = 0; i < num_pages; i++) {
3758                 struct page *p = eb->pages[i];
3759
3760                 clear_page_dirty_for_io(p);
3761                 set_page_writeback(p);
3762                 ret = submit_extent_page(rw, tree, wbc, p, offset >> 9,
3763                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3764                                          -1, end_bio_extent_buffer_writepage,
3765                                          0, epd->bio_flags, bio_flags);
3766                 epd->bio_flags = bio_flags;
3767                 if (ret) {
3768                         set_btree_ioerr(p);
3769                         end_page_writeback(p);
3770                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3771                                 end_extent_buffer_writeback(eb);
3772                         ret = -EIO;
3773                         break;
3774                 }
3775                 offset += PAGE_CACHE_SIZE;
3776                 update_nr_written(p, wbc, 1);
3777                 unlock_page(p);
3778         }
3779
3780         if (unlikely(ret)) {
3781                 for (; i < num_pages; i++) {
3782                         struct page *p = eb->pages[i];
3783                         clear_page_dirty_for_io(p);
3784                         unlock_page(p);
3785                 }
3786         }
3787
3788         return ret;
3789 }
3790
3791 int btree_write_cache_pages(struct address_space *mapping,
3792                                    struct writeback_control *wbc)
3793 {
3794         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3795         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3796         struct extent_buffer *eb, *prev_eb = NULL;
3797         struct extent_page_data epd = {
3798                 .bio = NULL,
3799                 .tree = tree,
3800                 .extent_locked = 0,
3801                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3802                 .bio_flags = 0,
3803         };
3804         int ret = 0;
3805         int done = 0;
3806         int nr_to_write_done = 0;
3807         struct pagevec pvec;
3808         int nr_pages;
3809         pgoff_t index;
3810         pgoff_t end;            /* Inclusive */
3811         int scanned = 0;
3812         int tag;
3813
3814         pagevec_init(&pvec, 0);
3815         if (wbc->range_cyclic) {
3816                 index = mapping->writeback_index; /* Start from prev offset */
3817                 end = -1;
3818         } else {
3819                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3820                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3821                 scanned = 1;
3822         }
3823         if (wbc->sync_mode == WB_SYNC_ALL)
3824                 tag = PAGECACHE_TAG_TOWRITE;
3825         else
3826                 tag = PAGECACHE_TAG_DIRTY;
3827 retry:
3828         if (wbc->sync_mode == WB_SYNC_ALL)
3829                 tag_pages_for_writeback(mapping, index, end);
3830         while (!done && !nr_to_write_done && (index <= end) &&
3831                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3832                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3833                 unsigned i;
3834
3835                 scanned = 1;
3836                 for (i = 0; i < nr_pages; i++) {
3837                         struct page *page = pvec.pages[i];
3838
3839                         if (!PagePrivate(page))
3840                                 continue;
3841
3842                         if (!wbc->range_cyclic && page->index > end) {
3843                                 done = 1;
3844                                 break;
3845                         }
3846
3847                         spin_lock(&mapping->private_lock);
3848                         if (!PagePrivate(page)) {
3849                                 spin_unlock(&mapping->private_lock);
3850                                 continue;
3851                         }
3852
3853                         eb = (struct extent_buffer *)page->private;
3854
3855                         /*
3856                          * Shouldn't happen and normally this would be a BUG_ON
3857                          * but no sense in crashing the users box for something
3858                          * we can survive anyway.
3859                          */
3860                         if (WARN_ON(!eb)) {
3861                                 spin_unlock(&mapping->private_lock);
3862                                 continue;
3863                         }
3864
3865                         if (eb == prev_eb) {
3866                                 spin_unlock(&mapping->private_lock);
3867                                 continue;
3868                         }
3869
3870                         ret = atomic_inc_not_zero(&eb->refs);
3871                         spin_unlock(&mapping->private_lock);
3872                         if (!ret)
3873                                 continue;
3874
3875                         prev_eb = eb;
3876                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3877                         if (!ret) {
3878                                 free_extent_buffer(eb);
3879                                 continue;
3880                         }
3881
3882                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3883                         if (ret) {
3884                                 done = 1;
3885                                 free_extent_buffer(eb);
3886                                 break;
3887                         }
3888                         free_extent_buffer(eb);
3889
3890                         /*
3891                          * the filesystem may choose to bump up nr_to_write.
3892                          * We have to make sure to honor the new nr_to_write
3893                          * at any time
3894                          */
3895                         nr_to_write_done = wbc->nr_to_write <= 0;
3896                 }
3897                 pagevec_release(&pvec);
3898                 cond_resched();
3899         }
3900         if (!scanned && !done) {
3901                 /*
3902                  * We hit the last page and there is more work to be done: wrap
3903                  * back to the start of the file
3904                  */
3905                 scanned = 1;
3906                 index = 0;
3907                 goto retry;
3908         }
3909         flush_write_bio(&epd);
3910         return ret;
3911 }
3912
3913 /**
3914  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3915  * @mapping: address space structure to write
3916  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3917  * @writepage: function called for each page
3918  * @data: data passed to writepage function
3919  *
3920  * If a page is already under I/O, write_cache_pages() skips it, even
3921  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3922  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3923  * and msync() need to guarantee that all the data which was dirty at the time
3924  * the call was made get new I/O started against them.  If wbc->sync_mode is
3925  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3926  * existing IO to complete.
3927  */
3928 static int extent_write_cache_pages(struct extent_io_tree *tree,
3929                              struct address_space *mapping,
3930                              struct writeback_control *wbc,
3931                              writepage_t writepage, void *data,
3932                              void (*flush_fn)(void *))
3933 {
3934         struct inode *inode = mapping->host;
3935         int ret = 0;
3936         int done = 0;
3937         int err = 0;
3938         int nr_to_write_done = 0;
3939         struct pagevec pvec;
3940         int nr_pages;
3941         pgoff_t index;
3942         pgoff_t end;            /* Inclusive */
3943         int scanned = 0;
3944         int tag;
3945
3946         /*
3947          * We have to hold onto the inode so that ordered extents can do their
3948          * work when the IO finishes.  The alternative to this is failing to add
3949          * an ordered extent if the igrab() fails there and that is a huge pain
3950          * to deal with, so instead just hold onto the inode throughout the
3951          * writepages operation.  If it fails here we are freeing up the inode
3952          * anyway and we'd rather not waste our time writing out stuff that is
3953          * going to be truncated anyway.
3954          */
3955         if (!igrab(inode))
3956                 return 0;
3957
3958         pagevec_init(&pvec, 0);
3959         if (wbc->range_cyclic) {
3960                 index = mapping->writeback_index; /* Start from prev offset */
3961                 end = -1;
3962         } else {
3963                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3964                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3965                 scanned = 1;
3966         }
3967         if (wbc->sync_mode == WB_SYNC_ALL)
3968                 tag = PAGECACHE_TAG_TOWRITE;
3969         else
3970                 tag = PAGECACHE_TAG_DIRTY;
3971 retry:
3972         if (wbc->sync_mode == WB_SYNC_ALL)
3973                 tag_pages_for_writeback(mapping, index, end);
3974         while (!done && !nr_to_write_done && (index <= end) &&
3975                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3976                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3977                 unsigned i;
3978
3979                 scanned = 1;
3980                 for (i = 0; i < nr_pages; i++) {
3981                         struct page *page = pvec.pages[i];
3982
3983                         /*
3984                          * At this point we hold neither mapping->tree_lock nor
3985                          * lock on the page itself: the page may be truncated or
3986                          * invalidated (changing page->mapping to NULL), or even
3987                          * swizzled back from swapper_space to tmpfs file
3988                          * mapping
3989                          */
3990                         if (!trylock_page(page)) {
3991                                 flush_fn(data);
3992                                 lock_page(page);
3993                         }
3994
3995                         if (unlikely(page->mapping != mapping)) {
3996                                 unlock_page(page);
3997                                 continue;
3998                         }
3999
4000                         if (!wbc->range_cyclic && page->index > end) {
4001                                 done = 1;
4002                                 unlock_page(page);
4003                                 continue;
4004                         }
4005
4006                         if (wbc->sync_mode != WB_SYNC_NONE) {
4007                                 if (PageWriteback(page))
4008                                         flush_fn(data);
4009                                 wait_on_page_writeback(page);
4010                         }
4011
4012                         if (PageWriteback(page) ||
4013                             !clear_page_dirty_for_io(page)) {
4014                                 unlock_page(page);
4015                                 continue;
4016                         }
4017
4018                         ret = (*writepage)(page, wbc, data);
4019
4020                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4021                                 unlock_page(page);
4022                                 ret = 0;
4023                         }
4024                         if (!err && ret < 0)
4025                                 err = ret;
4026
4027                         /*
4028                          * the filesystem may choose to bump up nr_to_write.
4029                          * We have to make sure to honor the new nr_to_write
4030                          * at any time
4031                          */
4032                         nr_to_write_done = wbc->nr_to_write <= 0;
4033                 }
4034                 pagevec_release(&pvec);
4035                 cond_resched();
4036         }
4037         if (!scanned && !done && !err) {
4038                 /*
4039                  * We hit the last page and there is more work to be done: wrap
4040                  * back to the start of the file
4041                  */
4042                 scanned = 1;
4043                 index = 0;
4044                 goto retry;
4045         }
4046         btrfs_add_delayed_iput(inode);
4047         return err;
4048 }
4049
4050 static void flush_epd_write_bio(struct extent_page_data *epd)
4051 {
4052         if (epd->bio) {
4053                 int rw = WRITE;
4054                 int ret;
4055
4056                 if (epd->sync_io)
4057                         rw = WRITE_SYNC;
4058
4059                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
4060                 BUG_ON(ret < 0); /* -ENOMEM */
4061                 epd->bio = NULL;
4062         }
4063 }
4064
4065 static noinline void flush_write_bio(void *data)
4066 {
4067         struct extent_page_data *epd = data;
4068         flush_epd_write_bio(epd);
4069 }
4070
4071 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4072                           get_extent_t *get_extent,
4073                           struct writeback_control *wbc)
4074 {
4075         int ret;
4076         struct extent_page_data epd = {
4077                 .bio = NULL,
4078                 .tree = tree,
4079                 .get_extent = get_extent,
4080                 .extent_locked = 0,
4081                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4082                 .bio_flags = 0,
4083         };
4084
4085         ret = __extent_writepage(page, wbc, &epd);
4086
4087         flush_epd_write_bio(&epd);
4088         return ret;
4089 }
4090
4091 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4092                               u64 start, u64 end, get_extent_t *get_extent,
4093                               int mode)
4094 {
4095         int ret = 0;
4096         struct address_space *mapping = inode->i_mapping;
4097         struct page *page;
4098         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
4099                 PAGE_CACHE_SHIFT;
4100
4101         struct extent_page_data epd = {
4102                 .bio = NULL,
4103                 .tree = tree,
4104                 .get_extent = get_extent,
4105                 .extent_locked = 1,
4106                 .sync_io = mode == WB_SYNC_ALL,
4107                 .bio_flags = 0,
4108         };
4109         struct writeback_control wbc_writepages = {
4110                 .sync_mode      = mode,
4111                 .nr_to_write    = nr_pages * 2,
4112                 .range_start    = start,
4113                 .range_end      = end + 1,
4114         };
4115
4116         while (start <= end) {
4117                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
4118                 if (clear_page_dirty_for_io(page))
4119                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4120                 else {
4121                         if (tree->ops && tree->ops->writepage_end_io_hook)
4122                                 tree->ops->writepage_end_io_hook(page, start,
4123                                                  start + PAGE_CACHE_SIZE - 1,
4124                                                  NULL, 1);
4125                         unlock_page(page);
4126                 }
4127                 page_cache_release(page);
4128                 start += PAGE_CACHE_SIZE;
4129         }
4130
4131         flush_epd_write_bio(&epd);
4132         return ret;
4133 }
4134
4135 int extent_writepages(struct extent_io_tree *tree,
4136                       struct address_space *mapping,
4137                       get_extent_t *get_extent,
4138                       struct writeback_control *wbc)
4139 {
4140         int ret = 0;
4141         struct extent_page_data epd = {
4142                 .bio = NULL,
4143                 .tree = tree,
4144                 .get_extent = get_extent,
4145                 .extent_locked = 0,
4146                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4147                 .bio_flags = 0,
4148         };
4149
4150         ret = extent_write_cache_pages(tree, mapping, wbc,
4151                                        __extent_writepage, &epd,
4152                                        flush_write_bio);
4153         flush_epd_write_bio(&epd);
4154         return ret;
4155 }
4156
4157 int extent_readpages(struct extent_io_tree *tree,
4158                      struct address_space *mapping,
4159                      struct list_head *pages, unsigned nr_pages,
4160                      get_extent_t get_extent)
4161 {
4162         struct bio *bio = NULL;
4163         unsigned page_idx;
4164         unsigned long bio_flags = 0;
4165         struct page *pagepool[16];
4166         struct page *page;
4167         struct extent_map *em_cached = NULL;
4168         int nr = 0;
4169
4170         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4171                 page = list_entry(pages->prev, struct page, lru);
4172
4173                 prefetchw(&page->flags);
4174                 list_del(&page->lru);
4175                 if (add_to_page_cache_lru(page, mapping,
4176                                         page->index, GFP_NOFS)) {
4177                         page_cache_release(page);
4178                         continue;
4179                 }
4180
4181                 pagepool[nr++] = page;
4182                 if (nr < ARRAY_SIZE(pagepool))
4183                         continue;
4184                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4185                                    &bio, 0, &bio_flags, READ);
4186                 nr = 0;
4187         }
4188         if (nr)
4189                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4190                                    &bio, 0, &bio_flags, READ);
4191
4192         if (em_cached)
4193                 free_extent_map(em_cached);
4194
4195         BUG_ON(!list_empty(pages));
4196         if (bio)
4197                 return submit_one_bio(READ, bio, 0, bio_flags);
4198         return 0;
4199 }
4200
4201 /*
4202  * basic invalidatepage code, this waits on any locked or writeback
4203  * ranges corresponding to the page, and then deletes any extent state
4204  * records from the tree
4205  */
4206 int extent_invalidatepage(struct extent_io_tree *tree,
4207                           struct page *page, unsigned long offset)
4208 {
4209         struct extent_state *cached_state = NULL;
4210         u64 start = page_offset(page);
4211         u64 end = start + PAGE_CACHE_SIZE - 1;
4212         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4213
4214         start += ALIGN(offset, blocksize);
4215         if (start > end)
4216                 return 0;
4217
4218         lock_extent_bits(tree, start, end, 0, &cached_state);
4219         wait_on_page_writeback(page);
4220         clear_extent_bit(tree, start, end,
4221                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4222                          EXTENT_DO_ACCOUNTING,
4223                          1, 1, &cached_state, GFP_NOFS);
4224         return 0;
4225 }
4226
4227 /*
4228  * a helper for releasepage, this tests for areas of the page that
4229  * are locked or under IO and drops the related state bits if it is safe
4230  * to drop the page.
4231  */
4232 static int try_release_extent_state(struct extent_map_tree *map,
4233                                     struct extent_io_tree *tree,
4234                                     struct page *page, gfp_t mask)
4235 {
4236         u64 start = page_offset(page);
4237         u64 end = start + PAGE_CACHE_SIZE - 1;
4238         int ret = 1;
4239
4240         if (test_range_bit(tree, start, end,
4241                            EXTENT_IOBITS, 0, NULL))
4242                 ret = 0;
4243         else {
4244                 if ((mask & GFP_NOFS) == GFP_NOFS)
4245                         mask = GFP_NOFS;
4246                 /*
4247                  * at this point we can safely clear everything except the
4248                  * locked bit and the nodatasum bit
4249                  */
4250                 ret = clear_extent_bit(tree, start, end,
4251                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4252                                  0, 0, NULL, mask);
4253
4254                 /* if clear_extent_bit failed for enomem reasons,
4255                  * we can't allow the release to continue.
4256                  */
4257                 if (ret < 0)
4258                         ret = 0;
4259                 else
4260                         ret = 1;
4261         }
4262         return ret;
4263 }
4264
4265 /*
4266  * a helper for releasepage.  As long as there are no locked extents
4267  * in the range corresponding to the page, both state records and extent
4268  * map records are removed
4269  */
4270 int try_release_extent_mapping(struct extent_map_tree *map,
4271                                struct extent_io_tree *tree, struct page *page,
4272                                gfp_t mask)
4273 {
4274         struct extent_map *em;
4275         u64 start = page_offset(page);
4276         u64 end = start + PAGE_CACHE_SIZE - 1;
4277
4278         if ((mask & __GFP_WAIT) &&
4279             page->mapping->host->i_size > 16 * 1024 * 1024) {
4280                 u64 len;
4281                 while (start <= end) {
4282                         len = end - start + 1;
4283                         write_lock(&map->lock);
4284                         em = lookup_extent_mapping(map, start, len);
4285                         if (!em) {
4286                                 write_unlock(&map->lock);
4287                                 break;
4288                         }
4289                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4290                             em->start != start) {
4291                                 write_unlock(&map->lock);
4292                                 free_extent_map(em);
4293                                 break;
4294                         }
4295                         if (!test_range_bit(tree, em->start,
4296                                             extent_map_end(em) - 1,
4297                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4298                                             0, NULL)) {
4299                                 remove_extent_mapping(map, em);
4300                                 /* once for the rb tree */
4301                                 free_extent_map(em);
4302                         }
4303                         start = extent_map_end(em);
4304                         write_unlock(&map->lock);
4305
4306                         /* once for us */
4307                         free_extent_map(em);
4308                 }
4309         }
4310         return try_release_extent_state(map, tree, page, mask);
4311 }
4312
4313 /*
4314  * helper function for fiemap, which doesn't want to see any holes.
4315  * This maps until we find something past 'last'
4316  */
4317 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4318                                                 u64 offset,
4319                                                 u64 last,
4320                                                 get_extent_t *get_extent)
4321 {
4322         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4323         struct extent_map *em;
4324         u64 len;
4325
4326         if (offset >= last)
4327                 return NULL;
4328
4329         while (1) {
4330                 len = last - offset;
4331                 if (len == 0)
4332                         break;
4333                 len = ALIGN(len, sectorsize);
4334                 em = get_extent(inode, NULL, 0, offset, len, 0);
4335                 if (IS_ERR_OR_NULL(em))
4336                         return em;
4337
4338                 /* if this isn't a hole return it */
4339                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4340                     em->block_start != EXTENT_MAP_HOLE) {
4341                         return em;
4342                 }
4343
4344                 /* this is a hole, advance to the next extent */
4345                 offset = extent_map_end(em);
4346                 free_extent_map(em);
4347                 if (offset >= last)
4348                         break;
4349         }
4350         return NULL;
4351 }
4352
4353 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4354                 __u64 start, __u64 len, get_extent_t *get_extent)
4355 {
4356         int ret = 0;
4357         u64 off = start;
4358         u64 max = start + len;
4359         u32 flags = 0;
4360         u32 found_type;
4361         u64 last;
4362         u64 last_for_get_extent = 0;
4363         u64 disko = 0;
4364         u64 isize = i_size_read(inode);
4365         struct btrfs_key found_key;
4366         struct extent_map *em = NULL;
4367         struct extent_state *cached_state = NULL;
4368         struct btrfs_path *path;
4369         struct btrfs_root *root = BTRFS_I(inode)->root;
4370         int end = 0;
4371         u64 em_start = 0;
4372         u64 em_len = 0;
4373         u64 em_end = 0;
4374
4375         if (len == 0)
4376                 return -EINVAL;
4377
4378         path = btrfs_alloc_path();
4379         if (!path)
4380                 return -ENOMEM;
4381         path->leave_spinning = 1;
4382
4383         start = round_down(start, BTRFS_I(inode)->root->sectorsize);
4384         len = round_up(max, BTRFS_I(inode)->root->sectorsize) - start;
4385
4386         /*
4387          * lookup the last file extent.  We're not using i_size here
4388          * because there might be preallocation past i_size
4389          */
4390         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode), -1,
4391                                        0);
4392         if (ret < 0) {
4393                 btrfs_free_path(path);
4394                 return ret;
4395         }
4396         WARN_ON(!ret);
4397         path->slots[0]--;
4398         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4399         found_type = found_key.type;
4400
4401         /* No extents, but there might be delalloc bits */
4402         if (found_key.objectid != btrfs_ino(inode) ||
4403             found_type != BTRFS_EXTENT_DATA_KEY) {
4404                 /* have to trust i_size as the end */
4405                 last = (u64)-1;
4406                 last_for_get_extent = isize;
4407         } else {
4408                 /*
4409                  * remember the start of the last extent.  There are a
4410                  * bunch of different factors that go into the length of the
4411                  * extent, so its much less complex to remember where it started
4412                  */
4413                 last = found_key.offset;
4414                 last_for_get_extent = last + 1;
4415         }
4416         btrfs_release_path(path);
4417
4418         /*
4419          * we might have some extents allocated but more delalloc past those
4420          * extents.  so, we trust isize unless the start of the last extent is
4421          * beyond isize
4422          */
4423         if (last < isize) {
4424                 last = (u64)-1;
4425                 last_for_get_extent = isize;
4426         }
4427
4428         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4429                          &cached_state);
4430
4431         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4432                                    get_extent);
4433         if (!em)
4434                 goto out;
4435         if (IS_ERR(em)) {
4436                 ret = PTR_ERR(em);
4437                 goto out;
4438         }
4439
4440         while (!end) {
4441                 u64 offset_in_extent = 0;
4442
4443                 /* break if the extent we found is outside the range */
4444                 if (em->start >= max || extent_map_end(em) < off)
4445                         break;
4446
4447                 /*
4448                  * get_extent may return an extent that starts before our
4449                  * requested range.  We have to make sure the ranges
4450                  * we return to fiemap always move forward and don't
4451                  * overlap, so adjust the offsets here
4452                  */
4453                 em_start = max(em->start, off);
4454
4455                 /*
4456                  * record the offset from the start of the extent
4457                  * for adjusting the disk offset below.  Only do this if the
4458                  * extent isn't compressed since our in ram offset may be past
4459                  * what we have actually allocated on disk.
4460                  */
4461                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4462                         offset_in_extent = em_start - em->start;
4463                 em_end = extent_map_end(em);
4464                 em_len = em_end - em_start;
4465                 disko = 0;
4466                 flags = 0;
4467
4468                 /*
4469                  * bump off for our next call to get_extent
4470                  */
4471                 off = extent_map_end(em);
4472                 if (off >= max)
4473                         end = 1;
4474
4475                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4476                         end = 1;
4477                         flags |= FIEMAP_EXTENT_LAST;
4478                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4479                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4480                                   FIEMAP_EXTENT_NOT_ALIGNED);
4481                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4482                         flags |= (FIEMAP_EXTENT_DELALLOC |
4483                                   FIEMAP_EXTENT_UNKNOWN);
4484                 } else if (fieinfo->fi_extents_max) {
4485                         u64 bytenr = em->block_start -
4486                                 (em->start - em->orig_start);
4487
4488                         disko = em->block_start + offset_in_extent;
4489
4490                         /*
4491                          * As btrfs supports shared space, this information
4492                          * can be exported to userspace tools via
4493                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4494                          * then we're just getting a count and we can skip the
4495                          * lookup stuff.
4496                          */
4497                         ret = btrfs_check_shared(NULL, root->fs_info,
4498                                                  root->objectid,
4499                                                  btrfs_ino(inode), bytenr);
4500                         if (ret < 0)
4501                                 goto out_free;
4502                         if (ret)
4503                                 flags |= FIEMAP_EXTENT_SHARED;
4504                         ret = 0;
4505                 }
4506                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4507                         flags |= FIEMAP_EXTENT_ENCODED;
4508                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4509                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4510
4511                 free_extent_map(em);
4512                 em = NULL;
4513                 if ((em_start >= last) || em_len == (u64)-1 ||
4514                    (last == (u64)-1 && isize <= em_end)) {
4515                         flags |= FIEMAP_EXTENT_LAST;
4516                         end = 1;
4517                 }
4518
4519                 /* now scan forward to see if this is really the last extent. */
4520                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4521                                            get_extent);
4522                 if (IS_ERR(em)) {
4523                         ret = PTR_ERR(em);
4524                         goto out;
4525                 }
4526                 if (!em) {
4527                         flags |= FIEMAP_EXTENT_LAST;
4528                         end = 1;
4529                 }
4530                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4531                                               em_len, flags);
4532                 if (ret) {
4533                         if (ret == 1)
4534                                 ret = 0;
4535                         goto out_free;
4536                 }
4537         }
4538 out_free:
4539         free_extent_map(em);
4540 out:
4541         btrfs_free_path(path);
4542         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4543                              &cached_state, GFP_NOFS);
4544         return ret;
4545 }
4546
4547 static void __free_extent_buffer(struct extent_buffer *eb)
4548 {
4549         btrfs_leak_debug_del(&eb->leak_list);
4550         kmem_cache_free(extent_buffer_cache, eb);
4551 }
4552
4553 int extent_buffer_under_io(struct extent_buffer *eb)
4554 {
4555         return (atomic_read(&eb->io_pages) ||
4556                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4557                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4558 }
4559
4560 /*
4561  * Helper for releasing extent buffer page.
4562  */
4563 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4564 {
4565         unsigned long index;
4566         struct page *page;
4567         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4568
4569         BUG_ON(extent_buffer_under_io(eb));
4570
4571         index = num_extent_pages(eb->start, eb->len);
4572         if (index == 0)
4573                 return;
4574
4575         do {
4576                 index--;
4577                 page = eb->pages[index];
4578                 if (!page)
4579                         continue;
4580                 if (mapped)
4581                         spin_lock(&page->mapping->private_lock);
4582                 /*
4583                  * We do this since we'll remove the pages after we've
4584                  * removed the eb from the radix tree, so we could race
4585                  * and have this page now attached to the new eb.  So
4586                  * only clear page_private if it's still connected to
4587                  * this eb.
4588                  */
4589                 if (PagePrivate(page) &&
4590                     page->private == (unsigned long)eb) {
4591                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4592                         BUG_ON(PageDirty(page));
4593                         BUG_ON(PageWriteback(page));
4594                         /*
4595                          * We need to make sure we haven't be attached
4596                          * to a new eb.
4597                          */
4598                         ClearPagePrivate(page);
4599                         set_page_private(page, 0);
4600                         /* One for the page private */
4601                         page_cache_release(page);
4602                 }
4603
4604                 if (mapped)
4605                         spin_unlock(&page->mapping->private_lock);
4606
4607                 /* One for when we alloced the page */
4608                 page_cache_release(page);
4609         } while (index != 0);
4610 }
4611
4612 /*
4613  * Helper for releasing the extent buffer.
4614  */
4615 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4616 {
4617         btrfs_release_extent_buffer_page(eb);
4618         __free_extent_buffer(eb);
4619 }
4620
4621 static struct extent_buffer *
4622 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4623                       unsigned long len)
4624 {
4625         struct extent_buffer *eb = NULL;
4626
4627         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4628         eb->start = start;
4629         eb->len = len;
4630         eb->fs_info = fs_info;
4631         eb->bflags = 0;
4632         rwlock_init(&eb->lock);
4633         atomic_set(&eb->write_locks, 0);
4634         atomic_set(&eb->read_locks, 0);
4635         atomic_set(&eb->blocking_readers, 0);
4636         atomic_set(&eb->blocking_writers, 0);
4637         atomic_set(&eb->spinning_readers, 0);
4638         atomic_set(&eb->spinning_writers, 0);
4639         eb->lock_nested = 0;
4640         init_waitqueue_head(&eb->write_lock_wq);
4641         init_waitqueue_head(&eb->read_lock_wq);
4642
4643         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4644
4645         spin_lock_init(&eb->refs_lock);
4646         atomic_set(&eb->refs, 1);
4647         atomic_set(&eb->io_pages, 0);
4648
4649         /*
4650          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4651          */
4652         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4653                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4654         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4655
4656         return eb;
4657 }
4658
4659 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4660 {
4661         unsigned long i;
4662         struct page *p;
4663         struct extent_buffer *new;
4664         unsigned long num_pages = num_extent_pages(src->start, src->len);
4665
4666         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4667         if (new == NULL)
4668                 return NULL;
4669
4670         for (i = 0; i < num_pages; i++) {
4671                 p = alloc_page(GFP_NOFS);
4672                 if (!p) {
4673                         btrfs_release_extent_buffer(new);
4674                         return NULL;
4675                 }
4676                 attach_extent_buffer_page(new, p);
4677                 WARN_ON(PageDirty(p));
4678                 SetPageUptodate(p);
4679                 new->pages[i] = p;
4680         }
4681
4682         copy_extent_buffer(new, src, 0, 0, src->len);
4683         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4684         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4685
4686         return new;
4687 }
4688
4689 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4690                                                 u64 start)
4691 {
4692         struct extent_buffer *eb;
4693         unsigned long len;
4694         unsigned long num_pages;
4695         unsigned long i;
4696
4697         if (!fs_info) {
4698                 /*
4699                  * Called only from tests that don't always have a fs_info
4700                  * available, but we know that nodesize is 4096
4701                  */
4702                 len = 4096;
4703         } else {
4704                 len = fs_info->tree_root->nodesize;
4705         }
4706         num_pages = num_extent_pages(0, len);
4707
4708         eb = __alloc_extent_buffer(fs_info, start, len);
4709         if (!eb)
4710                 return NULL;
4711
4712         for (i = 0; i < num_pages; i++) {
4713                 eb->pages[i] = alloc_page(GFP_NOFS);
4714                 if (!eb->pages[i])
4715                         goto err;
4716         }
4717         set_extent_buffer_uptodate(eb);
4718         btrfs_set_header_nritems(eb, 0);
4719         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4720
4721         return eb;
4722 err:
4723         for (; i > 0; i--)
4724                 __free_page(eb->pages[i - 1]);
4725         __free_extent_buffer(eb);
4726         return NULL;
4727 }
4728
4729 static void check_buffer_tree_ref(struct extent_buffer *eb)
4730 {
4731         int refs;
4732         /* the ref bit is tricky.  We have to make sure it is set
4733          * if we have the buffer dirty.   Otherwise the
4734          * code to free a buffer can end up dropping a dirty
4735          * page
4736          *
4737          * Once the ref bit is set, it won't go away while the
4738          * buffer is dirty or in writeback, and it also won't
4739          * go away while we have the reference count on the
4740          * eb bumped.
4741          *
4742          * We can't just set the ref bit without bumping the
4743          * ref on the eb because free_extent_buffer might
4744          * see the ref bit and try to clear it.  If this happens
4745          * free_extent_buffer might end up dropping our original
4746          * ref by mistake and freeing the page before we are able
4747          * to add one more ref.
4748          *
4749          * So bump the ref count first, then set the bit.  If someone
4750          * beat us to it, drop the ref we added.
4751          */
4752         refs = atomic_read(&eb->refs);
4753         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4754                 return;
4755
4756         spin_lock(&eb->refs_lock);
4757         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4758                 atomic_inc(&eb->refs);
4759         spin_unlock(&eb->refs_lock);
4760 }
4761
4762 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4763                 struct page *accessed)
4764 {
4765         unsigned long num_pages, i;
4766
4767         check_buffer_tree_ref(eb);
4768
4769         num_pages = num_extent_pages(eb->start, eb->len);
4770         for (i = 0; i < num_pages; i++) {
4771                 struct page *p = eb->pages[i];
4772
4773                 if (p != accessed)
4774                         mark_page_accessed(p);
4775         }
4776 }
4777
4778 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4779                                          u64 start)
4780 {
4781         struct extent_buffer *eb;
4782
4783         rcu_read_lock();
4784         eb = radix_tree_lookup(&fs_info->buffer_radix,
4785                                start >> PAGE_CACHE_SHIFT);
4786         if (eb && atomic_inc_not_zero(&eb->refs)) {
4787                 rcu_read_unlock();
4788                 /*
4789                  * Lock our eb's refs_lock to avoid races with
4790                  * free_extent_buffer. When we get our eb it might be flagged
4791                  * with EXTENT_BUFFER_STALE and another task running
4792                  * free_extent_buffer might have seen that flag set,
4793                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4794                  * writeback flags not set) and it's still in the tree (flag
4795                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4796                  * of decrementing the extent buffer's reference count twice.
4797                  * So here we could race and increment the eb's reference count,
4798                  * clear its stale flag, mark it as dirty and drop our reference
4799                  * before the other task finishes executing free_extent_buffer,
4800                  * which would later result in an attempt to free an extent
4801                  * buffer that is dirty.
4802                  */
4803                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4804                         spin_lock(&eb->refs_lock);
4805                         spin_unlock(&eb->refs_lock);
4806                 }
4807                 mark_extent_buffer_accessed(eb, NULL);
4808                 return eb;
4809         }
4810         rcu_read_unlock();
4811
4812         return NULL;
4813 }
4814
4815 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4816 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4817                                                u64 start)
4818 {
4819         struct extent_buffer *eb, *exists = NULL;
4820         int ret;
4821
4822         eb = find_extent_buffer(fs_info, start);
4823         if (eb)
4824                 return eb;
4825         eb = alloc_dummy_extent_buffer(fs_info, start);
4826         if (!eb)
4827                 return NULL;
4828         eb->fs_info = fs_info;
4829 again:
4830         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4831         if (ret)
4832                 goto free_eb;
4833         spin_lock(&fs_info->buffer_lock);
4834         ret = radix_tree_insert(&fs_info->buffer_radix,
4835                                 start >> PAGE_CACHE_SHIFT, eb);
4836         spin_unlock(&fs_info->buffer_lock);
4837         radix_tree_preload_end();
4838         if (ret == -EEXIST) {
4839                 exists = find_extent_buffer(fs_info, start);
4840                 if (exists)
4841                         goto free_eb;
4842                 else
4843                         goto again;
4844         }
4845         check_buffer_tree_ref(eb);
4846         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4847
4848         /*
4849          * We will free dummy extent buffer's if they come into
4850          * free_extent_buffer with a ref count of 2, but if we are using this we
4851          * want the buffers to stay in memory until we're done with them, so
4852          * bump the ref count again.
4853          */
4854         atomic_inc(&eb->refs);
4855         return eb;
4856 free_eb:
4857         btrfs_release_extent_buffer(eb);
4858         return exists;
4859 }
4860 #endif
4861
4862 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4863                                           u64 start)
4864 {
4865         unsigned long len = fs_info->tree_root->nodesize;
4866         unsigned long num_pages = num_extent_pages(start, len);
4867         unsigned long i;
4868         unsigned long index = start >> PAGE_CACHE_SHIFT;
4869         struct extent_buffer *eb;
4870         struct extent_buffer *exists = NULL;
4871         struct page *p;
4872         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4873         int uptodate = 1;
4874         int ret;
4875
4876         eb = find_extent_buffer(fs_info, start);
4877         if (eb)
4878                 return eb;
4879
4880         eb = __alloc_extent_buffer(fs_info, start, len);
4881         if (!eb)
4882                 return NULL;
4883
4884         for (i = 0; i < num_pages; i++, index++) {
4885                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4886                 if (!p)
4887                         goto free_eb;
4888
4889                 spin_lock(&mapping->private_lock);
4890                 if (PagePrivate(p)) {
4891                         /*
4892                          * We could have already allocated an eb for this page
4893                          * and attached one so lets see if we can get a ref on
4894                          * the existing eb, and if we can we know it's good and
4895                          * we can just return that one, else we know we can just
4896                          * overwrite page->private.
4897                          */
4898                         exists = (struct extent_buffer *)p->private;
4899                         if (atomic_inc_not_zero(&exists->refs)) {
4900                                 spin_unlock(&mapping->private_lock);
4901                                 unlock_page(p);
4902                                 page_cache_release(p);
4903                                 mark_extent_buffer_accessed(exists, p);
4904                                 goto free_eb;
4905                         }
4906                         exists = NULL;
4907
4908                         /*
4909                          * Do this so attach doesn't complain and we need to
4910                          * drop the ref the old guy had.
4911                          */
4912                         ClearPagePrivate(p);
4913                         WARN_ON(PageDirty(p));
4914                         page_cache_release(p);
4915                 }
4916                 attach_extent_buffer_page(eb, p);
4917                 spin_unlock(&mapping->private_lock);
4918                 WARN_ON(PageDirty(p));
4919                 eb->pages[i] = p;
4920                 if (!PageUptodate(p))
4921                         uptodate = 0;
4922
4923                 /*
4924                  * see below about how we avoid a nasty race with release page
4925                  * and why we unlock later
4926                  */
4927         }
4928         if (uptodate)
4929                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4930 again:
4931         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4932         if (ret)
4933                 goto free_eb;
4934
4935         spin_lock(&fs_info->buffer_lock);
4936         ret = radix_tree_insert(&fs_info->buffer_radix,
4937                                 start >> PAGE_CACHE_SHIFT, eb);
4938         spin_unlock(&fs_info->buffer_lock);
4939         radix_tree_preload_end();
4940         if (ret == -EEXIST) {
4941                 exists = find_extent_buffer(fs_info, start);
4942                 if (exists)
4943                         goto free_eb;
4944                 else
4945                         goto again;
4946         }
4947         /* add one reference for the tree */
4948         check_buffer_tree_ref(eb);
4949         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4950
4951         /*
4952          * there is a race where release page may have
4953          * tried to find this extent buffer in the radix
4954          * but failed.  It will tell the VM it is safe to
4955          * reclaim the, and it will clear the page private bit.
4956          * We must make sure to set the page private bit properly
4957          * after the extent buffer is in the radix tree so
4958          * it doesn't get lost
4959          */
4960         SetPageChecked(eb->pages[0]);
4961         for (i = 1; i < num_pages; i++) {
4962                 p = eb->pages[i];
4963                 ClearPageChecked(p);
4964                 unlock_page(p);
4965         }
4966         unlock_page(eb->pages[0]);
4967         return eb;
4968
4969 free_eb:
4970         WARN_ON(!atomic_dec_and_test(&eb->refs));
4971         for (i = 0; i < num_pages; i++) {
4972                 if (eb->pages[i])
4973                         unlock_page(eb->pages[i]);
4974         }
4975
4976         btrfs_release_extent_buffer(eb);
4977         return exists;
4978 }
4979
4980 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4981 {
4982         struct extent_buffer *eb =
4983                         container_of(head, struct extent_buffer, rcu_head);
4984
4985         __free_extent_buffer(eb);
4986 }
4987
4988 /* Expects to have eb->eb_lock already held */
4989 static int release_extent_buffer(struct extent_buffer *eb)
4990 {
4991         WARN_ON(atomic_read(&eb->refs) == 0);
4992         if (atomic_dec_and_test(&eb->refs)) {
4993                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
4994                         struct btrfs_fs_info *fs_info = eb->fs_info;
4995
4996                         spin_unlock(&eb->refs_lock);
4997
4998                         spin_lock(&fs_info->buffer_lock);
4999                         radix_tree_delete(&fs_info->buffer_radix,
5000                                           eb->start >> PAGE_CACHE_SHIFT);
5001                         spin_unlock(&fs_info->buffer_lock);
5002                 } else {
5003                         spin_unlock(&eb->refs_lock);
5004                 }
5005
5006                 /* Should be safe to release our pages at this point */
5007                 btrfs_release_extent_buffer_page(eb);
5008 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5009                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5010                         __free_extent_buffer(eb);
5011                         return 1;
5012                 }
5013 #endif
5014                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5015                 return 1;
5016         }
5017         spin_unlock(&eb->refs_lock);
5018
5019         return 0;
5020 }
5021
5022 void free_extent_buffer(struct extent_buffer *eb)
5023 {
5024         int refs;
5025         int old;
5026         if (!eb)
5027                 return;
5028
5029         while (1) {
5030                 refs = atomic_read(&eb->refs);
5031                 if (refs <= 3)
5032                         break;
5033                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5034                 if (old == refs)
5035                         return;
5036         }
5037
5038         spin_lock(&eb->refs_lock);
5039         if (atomic_read(&eb->refs) == 2 &&
5040             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5041                 atomic_dec(&eb->refs);
5042
5043         if (atomic_read(&eb->refs) == 2 &&
5044             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5045             !extent_buffer_under_io(eb) &&
5046             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5047                 atomic_dec(&eb->refs);
5048
5049         /*
5050          * I know this is terrible, but it's temporary until we stop tracking
5051          * the uptodate bits and such for the extent buffers.
5052          */
5053         release_extent_buffer(eb);
5054 }
5055
5056 void free_extent_buffer_stale(struct extent_buffer *eb)
5057 {
5058         if (!eb)
5059                 return;
5060
5061         spin_lock(&eb->refs_lock);
5062         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5063
5064         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5065             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5066                 atomic_dec(&eb->refs);
5067         release_extent_buffer(eb);
5068 }
5069
5070 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5071 {
5072         unsigned long i;
5073         unsigned long num_pages;
5074         struct page *page;
5075
5076         num_pages = num_extent_pages(eb->start, eb->len);
5077
5078         for (i = 0; i < num_pages; i++) {
5079                 page = eb->pages[i];
5080                 if (!PageDirty(page))
5081                         continue;
5082
5083                 lock_page(page);
5084                 WARN_ON(!PagePrivate(page));
5085
5086                 clear_page_dirty_for_io(page);
5087                 spin_lock_irq(&page->mapping->tree_lock);
5088                 if (!PageDirty(page)) {
5089                         radix_tree_tag_clear(&page->mapping->page_tree,
5090                                                 page_index(page),
5091                                                 PAGECACHE_TAG_DIRTY);
5092                 }
5093                 spin_unlock_irq(&page->mapping->tree_lock);
5094                 ClearPageError(page);
5095                 unlock_page(page);
5096         }
5097         WARN_ON(atomic_read(&eb->refs) == 0);
5098 }
5099
5100 int set_extent_buffer_dirty(struct extent_buffer *eb)
5101 {
5102         unsigned long i;
5103         unsigned long num_pages;
5104         int was_dirty = 0;
5105
5106         check_buffer_tree_ref(eb);
5107
5108         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5109
5110         num_pages = num_extent_pages(eb->start, eb->len);
5111         WARN_ON(atomic_read(&eb->refs) == 0);
5112         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5113
5114         for (i = 0; i < num_pages; i++)
5115                 set_page_dirty(eb->pages[i]);
5116         return was_dirty;
5117 }
5118
5119 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
5120 {
5121         unsigned long i;
5122         struct page *page;
5123         unsigned long num_pages;
5124
5125         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5126         num_pages = num_extent_pages(eb->start, eb->len);
5127         for (i = 0; i < num_pages; i++) {
5128                 page = eb->pages[i];
5129                 if (page)
5130                         ClearPageUptodate(page);
5131         }
5132         return 0;
5133 }
5134
5135 int set_extent_buffer_uptodate(struct extent_buffer *eb)
5136 {
5137         unsigned long i;
5138         struct page *page;
5139         unsigned long num_pages;
5140
5141         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5142         num_pages = num_extent_pages(eb->start, eb->len);
5143         for (i = 0; i < num_pages; i++) {
5144                 page = eb->pages[i];
5145                 SetPageUptodate(page);
5146         }
5147         return 0;
5148 }
5149
5150 int extent_buffer_uptodate(struct extent_buffer *eb)
5151 {
5152         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5153 }
5154
5155 int read_extent_buffer_pages(struct extent_io_tree *tree,
5156                              struct extent_buffer *eb, u64 start, int wait,
5157                              get_extent_t *get_extent, int mirror_num)
5158 {
5159         unsigned long i;
5160         unsigned long start_i;
5161         struct page *page;
5162         int err;
5163         int ret = 0;
5164         int locked_pages = 0;
5165         int all_uptodate = 1;
5166         unsigned long num_pages;
5167         unsigned long num_reads = 0;
5168         struct bio *bio = NULL;
5169         unsigned long bio_flags = 0;
5170
5171         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5172                 return 0;
5173
5174         if (start) {
5175                 WARN_ON(start < eb->start);
5176                 start_i = (start >> PAGE_CACHE_SHIFT) -
5177                         (eb->start >> PAGE_CACHE_SHIFT);
5178         } else {
5179                 start_i = 0;
5180         }
5181
5182         num_pages = num_extent_pages(eb->start, eb->len);
5183         for (i = start_i; i < num_pages; i++) {
5184                 page = eb->pages[i];
5185                 if (wait == WAIT_NONE) {
5186                         if (!trylock_page(page))
5187                                 goto unlock_exit;
5188                 } else {
5189                         lock_page(page);
5190                 }
5191                 locked_pages++;
5192                 if (!PageUptodate(page)) {
5193                         num_reads++;
5194                         all_uptodate = 0;
5195                 }
5196         }
5197         if (all_uptodate) {
5198                 if (start_i == 0)
5199                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5200                 goto unlock_exit;
5201         }
5202
5203         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5204         eb->read_mirror = 0;
5205         atomic_set(&eb->io_pages, num_reads);
5206         for (i = start_i; i < num_pages; i++) {
5207                 page = eb->pages[i];
5208                 if (!PageUptodate(page)) {
5209                         ClearPageError(page);
5210                         err = __extent_read_full_page(tree, page,
5211                                                       get_extent, &bio,
5212                                                       mirror_num, &bio_flags,
5213                                                       READ | REQ_META);
5214                         if (err)
5215                                 ret = err;
5216                 } else {
5217                         unlock_page(page);
5218                 }
5219         }
5220
5221         if (bio) {
5222                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
5223                                      bio_flags);
5224                 if (err)
5225                         return err;
5226         }
5227
5228         if (ret || wait != WAIT_COMPLETE)
5229                 return ret;
5230
5231         for (i = start_i; i < num_pages; i++) {
5232                 page = eb->pages[i];
5233                 wait_on_page_locked(page);
5234                 if (!PageUptodate(page))
5235                         ret = -EIO;
5236         }
5237
5238         return ret;
5239
5240 unlock_exit:
5241         i = start_i;
5242         while (locked_pages > 0) {
5243                 page = eb->pages[i];
5244                 i++;
5245                 unlock_page(page);
5246                 locked_pages--;
5247         }
5248         return ret;
5249 }
5250
5251 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5252                         unsigned long start,
5253                         unsigned long len)
5254 {
5255         size_t cur;
5256         size_t offset;
5257         struct page *page;
5258         char *kaddr;
5259         char *dst = (char *)dstv;
5260         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5261         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5262
5263         WARN_ON(start > eb->len);
5264         WARN_ON(start + len > eb->start + eb->len);
5265
5266         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5267
5268         while (len > 0) {
5269                 page = eb->pages[i];
5270
5271                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5272                 kaddr = page_address(page);
5273                 memcpy(dst, kaddr + offset, cur);
5274
5275                 dst += cur;
5276                 len -= cur;
5277                 offset = 0;
5278                 i++;
5279         }
5280 }
5281
5282 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5283                         unsigned long start,
5284                         unsigned long len)
5285 {
5286         size_t cur;
5287         size_t offset;
5288         struct page *page;
5289         char *kaddr;
5290         char __user *dst = (char __user *)dstv;
5291         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5292         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5293         int ret = 0;
5294
5295         WARN_ON(start > eb->len);
5296         WARN_ON(start + len > eb->start + eb->len);
5297
5298         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5299
5300         while (len > 0) {
5301                 page = eb->pages[i];
5302
5303                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5304                 kaddr = page_address(page);
5305                 if (copy_to_user(dst, kaddr + offset, cur)) {
5306                         ret = -EFAULT;
5307                         break;
5308                 }
5309
5310                 dst += cur;
5311                 len -= cur;
5312                 offset = 0;
5313                 i++;
5314         }
5315
5316         return ret;
5317 }
5318
5319 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5320                                unsigned long min_len, char **map,
5321                                unsigned long *map_start,
5322                                unsigned long *map_len)
5323 {
5324         size_t offset = start & (PAGE_CACHE_SIZE - 1);
5325         char *kaddr;
5326         struct page *p;
5327         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5328         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5329         unsigned long end_i = (start_offset + start + min_len - 1) >>
5330                 PAGE_CACHE_SHIFT;
5331
5332         if (i != end_i)
5333                 return -EINVAL;
5334
5335         if (i == 0) {
5336                 offset = start_offset;
5337                 *map_start = 0;
5338         } else {
5339                 offset = 0;
5340                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
5341         }
5342
5343         if (start + min_len > eb->len) {
5344                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
5345                        "wanted %lu %lu\n",
5346                        eb->start, eb->len, start, min_len);
5347                 return -EINVAL;
5348         }
5349
5350         p = eb->pages[i];
5351         kaddr = page_address(p);
5352         *map = kaddr + offset;
5353         *map_len = PAGE_CACHE_SIZE - offset;
5354         return 0;
5355 }
5356
5357 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5358                           unsigned long start,
5359                           unsigned long len)
5360 {
5361         size_t cur;
5362         size_t offset;
5363         struct page *page;
5364         char *kaddr;
5365         char *ptr = (char *)ptrv;
5366         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5367         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5368         int ret = 0;
5369
5370         WARN_ON(start > eb->len);
5371         WARN_ON(start + len > eb->start + eb->len);
5372
5373         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5374
5375         while (len > 0) {
5376                 page = eb->pages[i];
5377
5378                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5379
5380                 kaddr = page_address(page);
5381                 ret = memcmp(ptr, kaddr + offset, cur);
5382                 if (ret)
5383                         break;
5384
5385                 ptr += cur;
5386                 len -= cur;
5387                 offset = 0;
5388                 i++;
5389         }
5390         return ret;
5391 }
5392
5393 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5394                          unsigned long start, unsigned long len)
5395 {
5396         size_t cur;
5397         size_t offset;
5398         struct page *page;
5399         char *kaddr;
5400         char *src = (char *)srcv;
5401         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5402         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5403
5404         WARN_ON(start > eb->len);
5405         WARN_ON(start + len > eb->start + eb->len);
5406
5407         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5408
5409         while (len > 0) {
5410                 page = eb->pages[i];
5411                 WARN_ON(!PageUptodate(page));
5412
5413                 cur = min(len, PAGE_CACHE_SIZE - offset);
5414                 kaddr = page_address(page);
5415                 memcpy(kaddr + offset, src, cur);
5416
5417                 src += cur;
5418                 len -= cur;
5419                 offset = 0;
5420                 i++;
5421         }
5422 }
5423
5424 void memset_extent_buffer(struct extent_buffer *eb, char c,
5425                           unsigned long start, unsigned long len)
5426 {
5427         size_t cur;
5428         size_t offset;
5429         struct page *page;
5430         char *kaddr;
5431         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5432         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5433
5434         WARN_ON(start > eb->len);
5435         WARN_ON(start + len > eb->start + eb->len);
5436
5437         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5438
5439         while (len > 0) {
5440                 page = eb->pages[i];
5441                 WARN_ON(!PageUptodate(page));
5442
5443                 cur = min(len, PAGE_CACHE_SIZE - offset);
5444                 kaddr = page_address(page);
5445                 memset(kaddr + offset, c, cur);
5446
5447                 len -= cur;
5448                 offset = 0;
5449                 i++;
5450         }
5451 }
5452
5453 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5454                         unsigned long dst_offset, unsigned long src_offset,
5455                         unsigned long len)
5456 {
5457         u64 dst_len = dst->len;
5458         size_t cur;
5459         size_t offset;
5460         struct page *page;
5461         char *kaddr;
5462         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5463         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5464
5465         WARN_ON(src->len != dst_len);
5466
5467         offset = (start_offset + dst_offset) &
5468                 (PAGE_CACHE_SIZE - 1);
5469
5470         while (len > 0) {
5471                 page = dst->pages[i];
5472                 WARN_ON(!PageUptodate(page));
5473
5474                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5475
5476                 kaddr = page_address(page);
5477                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5478
5479                 src_offset += cur;
5480                 len -= cur;
5481                 offset = 0;
5482                 i++;
5483         }
5484 }
5485
5486 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5487 {
5488         unsigned long distance = (src > dst) ? src - dst : dst - src;
5489         return distance < len;
5490 }
5491
5492 static void copy_pages(struct page *dst_page, struct page *src_page,
5493                        unsigned long dst_off, unsigned long src_off,
5494                        unsigned long len)
5495 {
5496         char *dst_kaddr = page_address(dst_page);
5497         char *src_kaddr;
5498         int must_memmove = 0;
5499
5500         if (dst_page != src_page) {
5501                 src_kaddr = page_address(src_page);
5502         } else {
5503                 src_kaddr = dst_kaddr;
5504                 if (areas_overlap(src_off, dst_off, len))
5505                         must_memmove = 1;
5506         }
5507
5508         if (must_memmove)
5509                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5510         else
5511                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5512 }
5513
5514 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5515                            unsigned long src_offset, unsigned long len)
5516 {
5517         size_t cur;
5518         size_t dst_off_in_page;
5519         size_t src_off_in_page;
5520         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5521         unsigned long dst_i;
5522         unsigned long src_i;
5523
5524         if (src_offset + len > dst->len) {
5525                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5526                        "len %lu dst len %lu\n", src_offset, len, dst->len);
5527                 BUG_ON(1);
5528         }
5529         if (dst_offset + len > dst->len) {
5530                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5531                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
5532                 BUG_ON(1);
5533         }
5534
5535         while (len > 0) {
5536                 dst_off_in_page = (start_offset + dst_offset) &
5537                         (PAGE_CACHE_SIZE - 1);
5538                 src_off_in_page = (start_offset + src_offset) &
5539                         (PAGE_CACHE_SIZE - 1);
5540
5541                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5542                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5543
5544                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5545                                                src_off_in_page));
5546                 cur = min_t(unsigned long, cur,
5547                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5548
5549                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5550                            dst_off_in_page, src_off_in_page, cur);
5551
5552                 src_offset += cur;
5553                 dst_offset += cur;
5554                 len -= cur;
5555         }
5556 }
5557
5558 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5559                            unsigned long src_offset, unsigned long len)
5560 {
5561         size_t cur;
5562         size_t dst_off_in_page;
5563         size_t src_off_in_page;
5564         unsigned long dst_end = dst_offset + len - 1;
5565         unsigned long src_end = src_offset + len - 1;
5566         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5567         unsigned long dst_i;
5568         unsigned long src_i;
5569
5570         if (src_offset + len > dst->len) {
5571                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5572                        "len %lu len %lu\n", src_offset, len, dst->len);
5573                 BUG_ON(1);
5574         }
5575         if (dst_offset + len > dst->len) {
5576                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5577                        "len %lu len %lu\n", dst_offset, len, dst->len);
5578                 BUG_ON(1);
5579         }
5580         if (dst_offset < src_offset) {
5581                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5582                 return;
5583         }
5584         while (len > 0) {
5585                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5586                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5587
5588                 dst_off_in_page = (start_offset + dst_end) &
5589                         (PAGE_CACHE_SIZE - 1);
5590                 src_off_in_page = (start_offset + src_end) &
5591                         (PAGE_CACHE_SIZE - 1);
5592
5593                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5594                 cur = min(cur, dst_off_in_page + 1);
5595                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5596                            dst_off_in_page - cur + 1,
5597                            src_off_in_page - cur + 1, cur);
5598
5599                 dst_end -= cur;
5600                 src_end -= cur;
5601                 len -= cur;
5602         }
5603 }
5604
5605 int try_release_extent_buffer(struct page *page)
5606 {
5607         struct extent_buffer *eb;
5608
5609         /*
5610          * We need to make sure noboody is attaching this page to an eb right
5611          * now.
5612          */
5613         spin_lock(&page->mapping->private_lock);
5614         if (!PagePrivate(page)) {
5615                 spin_unlock(&page->mapping->private_lock);
5616                 return 1;
5617         }
5618
5619         eb = (struct extent_buffer *)page->private;
5620         BUG_ON(!eb);
5621
5622         /*
5623          * This is a little awful but should be ok, we need to make sure that
5624          * the eb doesn't disappear out from under us while we're looking at
5625          * this page.
5626          */
5627         spin_lock(&eb->refs_lock);
5628         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5629                 spin_unlock(&eb->refs_lock);
5630                 spin_unlock(&page->mapping->private_lock);
5631                 return 0;
5632         }
5633         spin_unlock(&page->mapping->private_lock);
5634
5635         /*
5636          * If tree ref isn't set then we know the ref on this eb is a real ref,
5637          * so just return, this page will likely be freed soon anyway.
5638          */
5639         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5640                 spin_unlock(&eb->refs_lock);
5641                 return 0;
5642         }
5643
5644         return release_extent_buffer(eb);
5645 }