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                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
 100                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
 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 void add_extent_changeset(struct extent_state *state, unsigned bits,
 135                                  struct extent_changeset *changeset,
 136                                  int set)
 137 {
 138         int ret;
 139
 140         if (!changeset)
 141                 return;
 142         if (set && (state->state & bits) == bits)
 143                 return;
 144         if (!set && (state->state & bits) == 0)
 145                 return;
 146         changeset->bytes_changed += state->end - state->start + 1;
 147         ret = ulist_add(changeset->range_changed, state->start, state->end,
 148                         GFP_ATOMIC);
 149         /* ENOMEM */
 150         BUG_ON(ret < 0);
 151 }
 152
 153 static noinline void flush_write_bio(void *data);
 154 static inline struct btrfs_fs_info *
 155 tree_fs_info(struct extent_io_tree *tree)
 156 {
 157         if (!tree->mapping)
 158                 return NULL;
 159         return btrfs_sb(tree->mapping->host->i_sb);
 160 }
 161
 162 int __init extent_io_init(void)
 163 {
 164         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 165                         sizeof(struct extent_state), 0,
 166                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 167         if (!extent_state_cache)
 168                 return -ENOMEM;
 169
 170         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 171                         sizeof(struct extent_buffer), 0,
 172                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 173         if (!extent_buffer_cache)
 174                 goto free_state_cache;
 175
 176         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 177                                      offsetof(struct btrfs_io_bio, bio));
 178         if (!btrfs_bioset)
 179                 goto free_buffer_cache;
 180
 181         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 182                 goto free_bioset;
 183
 184         return 0;
 185
 186 free_bioset:
 187         bioset_free(btrfs_bioset);
 188         btrfs_bioset = NULL;
 189
 190 free_buffer_cache:
 191         kmem_cache_destroy(extent_buffer_cache);
 192         extent_buffer_cache = NULL;
 193
 194 free_state_cache:
 195         kmem_cache_destroy(extent_state_cache);
 196         extent_state_cache = NULL;
 197         return -ENOMEM;
 198 }
 199
 200 void extent_io_exit(void)
 201 {
 202         btrfs_leak_debug_check();
 203
 204         /*
 205          * Make sure all delayed rcu free are flushed before we
 206          * destroy caches.
 207          */
 208         rcu_barrier();
 209         if (extent_state_cache)
 210                 kmem_cache_destroy(extent_state_cache);
 211         if (extent_buffer_cache)
 212                 kmem_cache_destroy(extent_buffer_cache);
 213         if (btrfs_bioset)
 214                 bioset_free(btrfs_bioset);
 215 }
 216
 217 void extent_io_tree_init(struct extent_io_tree *tree,
 218                          struct address_space *mapping)
 219 {
 220         tree->state = RB_ROOT;
 221         tree->ops = NULL;
 222         tree->dirty_bytes = 0;
 223         spin_lock_init(&tree->lock);
 224         tree->mapping = mapping;
 225 }
 226
 227 static struct extent_state *alloc_extent_state(gfp_t mask)
 228 {
 229         struct extent_state *state;
 230
 231         state = kmem_cache_alloc(extent_state_cache, mask);
 232         if (!state)
 233                 return state;
 234         state->state = 0;
 235         state->private = 0;
 236         RB_CLEAR_NODE(&state->rb_node);
 237         btrfs_leak_debug_add(&state->leak_list, &states);
 238         atomic_set(&state->refs, 1);
 239         init_waitqueue_head(&state->wq);
 240         trace_alloc_extent_state(state, mask, _RET_IP_);
 241         return state;
 242 }
 243
 244 void free_extent_state(struct extent_state *state)
 245 {
 246         if (!state)
 247                 return;
 248         if (atomic_dec_and_test(&state->refs)) {
 249                 WARN_ON(extent_state_in_tree(state));
 250                 btrfs_leak_debug_del(&state->leak_list);
 251                 trace_free_extent_state(state, _RET_IP_);
 252                 kmem_cache_free(extent_state_cache, state);
 253         }
 254 }
 255
 256 static struct rb_node *tree_insert(struct rb_root *root,
 257                                    struct rb_node *search_start,
 258                                    u64 offset,
 259                                    struct rb_node *node,
 260                                    struct rb_node ***p_in,
 261                                    struct rb_node **parent_in)
 262 {
 263         struct rb_node **p;
 264         struct rb_node *parent = NULL;
 265         struct tree_entry *entry;
 266
 267         if (p_in && parent_in) {
 268                 p = *p_in;
 269                 parent = *parent_in;
 270                 goto do_insert;
 271         }
 272
 273         p = search_start ? &search_start : &root->rb_node;
 274         while (*p) {
 275                 parent = *p;
 276                 entry = rb_entry(parent, struct tree_entry, rb_node);
 277
 278                 if (offset < entry->start)
 279                         p = &(*p)->rb_left;
 280                 else if (offset > entry->end)
 281                         p = &(*p)->rb_right;
 282                 else
 283                         return parent;
 284         }
 285
 286 do_insert:
 287         rb_link_node(node, parent, p);
 288         rb_insert_color(node, root);
 289         return NULL;
 290 }
 291
 292 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 293                                       struct rb_node **prev_ret,
 294                                       struct rb_node **next_ret,
 295                                       struct rb_node ***p_ret,
 296                                       struct rb_node **parent_ret)
 297 {
 298         struct rb_root *root = &tree->state;
 299         struct rb_node **n = &root->rb_node;
 300         struct rb_node *prev = NULL;
 301         struct rb_node *orig_prev = NULL;
 302         struct tree_entry *entry;
 303         struct tree_entry *prev_entry = NULL;
 304
 305         while (*n) {
 306                 prev = *n;
 307                 entry = rb_entry(prev, struct tree_entry, rb_node);
 308                 prev_entry = entry;
 309
 310                 if (offset < entry->start)
 311                         n = &(*n)->rb_left;
 312                 else if (offset > entry->end)
 313                         n = &(*n)->rb_right;
 314                 else
 315                         return *n;
 316         }
 317
 318         if (p_ret)
 319                 *p_ret = n;
 320         if (parent_ret)
 321                 *parent_ret = prev;
 322
 323         if (prev_ret) {
 324                 orig_prev = prev;
 325                 while (prev && offset > prev_entry->end) {
 326                         prev = rb_next(prev);
 327                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 328                 }
 329                 *prev_ret = prev;
 330                 prev = orig_prev;
 331         }
 332
 333         if (next_ret) {
 334                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 335                 while (prev && offset < prev_entry->start) {
 336                         prev = rb_prev(prev);
 337                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 338                 }
 339                 *next_ret = prev;
 340         }
 341         return NULL;
 342 }
 343
 344 static inline struct rb_node *
 345 tree_search_for_insert(struct extent_io_tree *tree,
 346                        u64 offset,
 347                        struct rb_node ***p_ret,
 348                        struct rb_node **parent_ret)
 349 {
 350         struct rb_node *prev = NULL;
 351         struct rb_node *ret;
 352
 353         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 354         if (!ret)
 355                 return prev;
 356         return ret;
 357 }
 358
 359 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 360                                           u64 offset)
 361 {
 362         return tree_search_for_insert(tree, offset, NULL, NULL);
 363 }
 364
 365 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 366                      struct extent_state *other)
 367 {
 368         if (tree->ops && tree->ops->merge_extent_hook)
 369                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 370                                              other);
 371 }
 372
 373 /*
 374  * utility function to look for merge candidates inside a given range.
 375  * Any extents with matching state are merged together into a single
 376  * extent in the tree.  Extents with EXTENT_IO in their state field
 377  * are not merged because the end_io handlers need to be able to do
 378  * operations on them without sleeping (or doing allocations/splits).
 379  *
 380  * This should be called with the tree lock held.
 381  */
 382 static void merge_state(struct extent_io_tree *tree,
 383                         struct extent_state *state)
 384 {
 385         struct extent_state *other;
 386         struct rb_node *other_node;
 387
 388         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 389                 return;
 390
 391         other_node = rb_prev(&state->rb_node);
 392         if (other_node) {
 393                 other = rb_entry(other_node, struct extent_state, rb_node);
 394                 if (other->end == state->start - 1 &&
 395                     other->state == state->state) {
 396                         merge_cb(tree, state, other);
 397                         state->start = other->start;
 398                         rb_erase(&other->rb_node, &tree->state);
 399                         RB_CLEAR_NODE(&other->rb_node);
 400                         free_extent_state(other);
 401                 }
 402         }
 403         other_node = rb_next(&state->rb_node);
 404         if (other_node) {
 405                 other = rb_entry(other_node, struct extent_state, rb_node);
 406                 if (other->start == state->end + 1 &&
 407                     other->state == state->state) {
 408                         merge_cb(tree, state, other);
 409                         state->end = other->end;
 410                         rb_erase(&other->rb_node, &tree->state);
 411                         RB_CLEAR_NODE(&other->rb_node);
 412                         free_extent_state(other);
 413                 }
 414         }
 415 }
 416
 417 static void set_state_cb(struct extent_io_tree *tree,
 418                          struct extent_state *state, unsigned *bits)
 419 {
 420         if (tree->ops && tree->ops->set_bit_hook)
 421                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 422 }
 423
 424 static void clear_state_cb(struct extent_io_tree *tree,
 425                            struct extent_state *state, unsigned *bits)
 426 {
 427         if (tree->ops && tree->ops->clear_bit_hook)
 428                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 429 }
 430
 431 static void set_state_bits(struct extent_io_tree *tree,
 432                            struct extent_state *state, unsigned *bits,
 433                            struct extent_changeset *changeset);
 434
 435 /*
 436  * insert an extent_state struct into the tree.  'bits' are set on the
 437  * struct before it is inserted.
 438  *
 439  * This may return -EEXIST if the extent is already there, in which case the
 440  * state struct is freed.
 441  *
 442  * The tree lock is not taken internally.  This is a utility function and
 443  * probably isn't what you want to call (see set/clear_extent_bit).
 444  */
 445 static int insert_state(struct extent_io_tree *tree,
 446                         struct extent_state *state, u64 start, u64 end,
 447                         struct rb_node ***p,
 448                         struct rb_node **parent,
 449                         unsigned *bits, struct extent_changeset *changeset)
 450 {
 451         struct rb_node *node;
 452
 453         if (end < start)
 454                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 455                        end, start);
 456         state->start = start;
 457         state->end = end;
 458
 459         set_state_bits(tree, state, bits, changeset);
 460
 461         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 462         if (node) {
 463                 struct extent_state *found;
 464                 found = rb_entry(node, struct extent_state, rb_node);
 465                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
 466                        "%llu %llu\n",
 467                        found->start, found->end, start, end);
 468                 return -EEXIST;
 469         }
 470         merge_state(tree, state);
 471         return 0;
 472 }
 473
 474 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 475                      u64 split)
 476 {
 477         if (tree->ops && tree->ops->split_extent_hook)
 478                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 479 }
 480
 481 /*
 482  * split a given extent state struct in two, inserting the preallocated
 483  * struct 'prealloc' as the newly created second half.  'split' indicates an
 484  * offset inside 'orig' where it should be split.
 485  *
 486  * Before calling,
 487  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 488  * are two extent state structs in the tree:
 489  * prealloc: [orig->start, split - 1]
 490  * orig: [ split, orig->end ]
 491  *
 492  * The tree locks are not taken by this function. They need to be held
 493  * by the caller.
 494  */
 495 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 496                        struct extent_state *prealloc, u64 split)
 497 {
 498         struct rb_node *node;
 499
 500         split_cb(tree, orig, split);
 501
 502         prealloc->start = orig->start;
 503         prealloc->end = split - 1;
 504         prealloc->state = orig->state;
 505         orig->start = split;
 506
 507         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 508                            &prealloc->rb_node, NULL, NULL);
 509         if (node) {
 510                 free_extent_state(prealloc);
 511                 return -EEXIST;
 512         }
 513         return 0;
 514 }
 515
 516 static struct extent_state *next_state(struct extent_state *state)
 517 {
 518         struct rb_node *next = rb_next(&state->rb_node);
 519         if (next)
 520                 return rb_entry(next, struct extent_state, rb_node);
 521         else
 522                 return NULL;
 523 }
 524
 525 /*
 526  * utility function to clear some bits in an extent state struct.
 527  * it will optionally wake up any one waiting on this state (wake == 1).
 528  *
 529  * If no bits are set on the state struct after clearing things, the
 530  * struct is freed and removed from the tree
 531  */
 532 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 533                                             struct extent_state *state,
 534                                             unsigned *bits, int wake,
 535                                             struct extent_changeset *changeset)
 536 {
 537         struct extent_state *next;
 538         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 539
 540         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 541                 u64 range = state->end - state->start + 1;
 542                 WARN_ON(range > tree->dirty_bytes);
 543                 tree->dirty_bytes -= range;
 544         }
 545         clear_state_cb(tree, state, bits);
 546         add_extent_changeset(state, bits_to_clear, changeset, 0);
 547         state->state &= ~bits_to_clear;
 548         if (wake)
 549                 wake_up(&state->wq);
 550         if (state->state == 0) {
 551                 next = next_state(state);
 552                 if (extent_state_in_tree(state)) {
 553                         rb_erase(&state->rb_node, &tree->state);
 554                         RB_CLEAR_NODE(&state->rb_node);
 555                         free_extent_state(state);
 556                 } else {
 557                         WARN_ON(1);
 558                 }
 559         } else {
 560                 merge_state(tree, state);
 561                 next = next_state(state);
 562         }
 563         return next;
 564 }
 565
 566 static struct extent_state *
 567 alloc_extent_state_atomic(struct extent_state *prealloc)
 568 {
 569         if (!prealloc)
 570                 prealloc = alloc_extent_state(GFP_ATOMIC);
 571
 572         return prealloc;
 573 }
 574
 575 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 576 {
 577         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 578                     "Extent tree was modified by another "
 579                     "thread while locked.");
 580 }
 581
 582 /*
 583  * clear some bits on a range in the tree.  This may require splitting
 584  * or inserting elements in the tree, so the gfp mask is used to
 585  * indicate which allocations or sleeping are allowed.
 586  *
 587  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 588  * the given range from the tree regardless of state (ie for truncate).
 589  *
 590  * the range [start, end] is inclusive.
 591  *
 592  * This takes the tree lock, and returns 0 on success and < 0 on error.
 593  */
 594 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 595                               unsigned bits, int wake, int delete,
 596                               struct extent_state **cached_state,
 597                               gfp_t mask, struct extent_changeset *changeset)
 598 {
 599         struct extent_state *state;
 600         struct extent_state *cached;
 601         struct extent_state *prealloc = NULL;
 602         struct rb_node *node;
 603         u64 last_end;
 604         int err;
 605         int clear = 0;
 606
 607         btrfs_debug_check_extent_io_range(tree, start, end);
 608
 609         if (bits & EXTENT_DELALLOC)
 610                 bits |= EXTENT_NORESERVE;
 611
 612         if (delete)
 613                 bits |= ~EXTENT_CTLBITS;
 614         bits |= EXTENT_FIRST_DELALLOC;
 615
 616         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 617                 clear = 1;
 618 again:
 619         if (!prealloc && gfpflags_allow_blocking(mask)) {
 620                 /*
 621                  * Don't care for allocation failure here because we might end
 622                  * up not needing the pre-allocated extent state at all, which
 623                  * is the case if we only have in the tree extent states that
 624                  * cover our input range and don't cover too any other range.
 625                  * If we end up needing a new extent state we allocate it later.
 626                  */
 627                 prealloc = alloc_extent_state(mask);
 628         }
 629
 630         spin_lock(&tree->lock);
 631         if (cached_state) {
 632                 cached = *cached_state;
 633
 634                 if (clear) {
 635                         *cached_state = NULL;
 636                         cached_state = NULL;
 637                 }
 638
 639                 if (cached && extent_state_in_tree(cached) &&
 640                     cached->start <= start && cached->end > start) {
 641                         if (clear)
 642                                 atomic_dec(&cached->refs);
 643                         state = cached;
 644                         goto hit_next;
 645                 }
 646                 if (clear)
 647                         free_extent_state(cached);
 648         }
 649         /*
 650          * this search will find the extents that end after
 651          * our range starts
 652          */
 653         node = tree_search(tree, start);
 654         if (!node)
 655                 goto out;
 656         state = rb_entry(node, struct extent_state, rb_node);
 657 hit_next:
 658         if (state->start > end)
 659                 goto out;
 660         WARN_ON(state->end < start);
 661         last_end = state->end;
 662
 663         /* the state doesn't have the wanted bits, go ahead */
 664         if (!(state->state & bits)) {
 665                 state = next_state(state);
 666                 goto next;
 667         }
 668
 669         /*
 670          *     | ---- desired range ---- |
 671          *  | state | or
 672          *  | ------------- state -------------- |
 673          *
 674          * We need to split the extent we found, and may flip
 675          * bits on second half.
 676          *
 677          * If the extent we found extends past our range, we
 678          * just split and search again.  It'll get split again
 679          * the next time though.
 680          *
 681          * If the extent we found is inside our range, we clear
 682          * the desired bit on it.
 683          */
 684
 685         if (state->start < start) {
 686                 prealloc = alloc_extent_state_atomic(prealloc);
 687                 BUG_ON(!prealloc);
 688                 err = split_state(tree, state, prealloc, start);
 689                 if (err)
 690                         extent_io_tree_panic(tree, err);
 691
 692                 prealloc = NULL;
 693                 if (err)
 694                         goto out;
 695                 if (state->end <= end) {
 696                         state = clear_state_bit(tree, state, &bits, wake,
 697                                                 changeset);
 698                         goto next;
 699                 }
 700                 goto search_again;
 701         }
 702         /*
 703          * | ---- desired range ---- |
 704          *                        | state |
 705          * We need to split the extent, and clear the bit
 706          * on the first half
 707          */
 708         if (state->start <= end && state->end > end) {
 709                 prealloc = alloc_extent_state_atomic(prealloc);
 710                 BUG_ON(!prealloc);
 711                 err = split_state(tree, state, prealloc, end + 1);
 712                 if (err)
 713                         extent_io_tree_panic(tree, err);
 714
 715                 if (wake)
 716                         wake_up(&state->wq);
 717
 718                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 719
 720                 prealloc = NULL;
 721                 goto out;
 722         }
 723
 724         state = clear_state_bit(tree, state, &bits, wake, changeset);
 725 next:
 726         if (last_end == (u64)-1)
 727                 goto out;
 728         start = last_end + 1;
 729         if (start <= end && state && !need_resched())
 730                 goto hit_next;
 731         goto search_again;
 732
 733 out:
 734         spin_unlock(&tree->lock);
 735         if (prealloc)
 736                 free_extent_state(prealloc);
 737
 738         return 0;
 739
 740 search_again:
 741         if (start > end)
 742                 goto out;
 743         spin_unlock(&tree->lock);
 744         if (gfpflags_allow_blocking(mask))
 745                 cond_resched();
 746         goto again;
 747 }
 748
 749 static void wait_on_state(struct extent_io_tree *tree,
 750                           struct extent_state *state)
 751                 __releases(tree->lock)
 752                 __acquires(tree->lock)
 753 {
 754         DEFINE_WAIT(wait);
 755         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 756         spin_unlock(&tree->lock);
 757         schedule();
 758         spin_lock(&tree->lock);
 759         finish_wait(&state->wq, &wait);
 760 }
 761
 762 /*
 763  * waits for one or more bits to clear on a range in the state tree.
 764  * The range [start, end] is inclusive.
 765  * The tree lock is taken by this function
 766  */
 767 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 768                             unsigned long bits)
 769 {
 770         struct extent_state *state;
 771         struct rb_node *node;
 772
 773         btrfs_debug_check_extent_io_range(tree, start, end);
 774
 775         spin_lock(&tree->lock);
 776 again:
 777         while (1) {
 778                 /*
 779                  * this search will find all the extents that end after
 780                  * our range starts
 781                  */
 782                 node = tree_search(tree, start);
 783 process_node:
 784                 if (!node)
 785                         break;
 786
 787                 state = rb_entry(node, struct extent_state, rb_node);
 788
 789                 if (state->start > end)
 790                         goto out;
 791
 792                 if (state->state & bits) {
 793                         start = state->start;
 794                         atomic_inc(&state->refs);
 795                         wait_on_state(tree, state);
 796                         free_extent_state(state);
 797                         goto again;
 798                 }
 799                 start = state->end + 1;
 800
 801                 if (start > end)
 802                         break;
 803
 804                 if (!cond_resched_lock(&tree->lock)) {
 805                         node = rb_next(node);
 806                         goto process_node;
 807                 }
 808         }
 809 out:
 810         spin_unlock(&tree->lock);
 811 }
 812
 813 static void set_state_bits(struct extent_io_tree *tree,
 814                            struct extent_state *state,
 815                            unsigned *bits, struct extent_changeset *changeset)
 816 {
 817         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 818
 819         set_state_cb(tree, state, bits);
 820         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 821                 u64 range = state->end - state->start + 1;
 822                 tree->dirty_bytes += range;
 823         }
 824         add_extent_changeset(state, bits_to_set, changeset, 1);
 825         state->state |= bits_to_set;
 826 }
 827
 828 static void cache_state_if_flags(struct extent_state *state,
 829                                  struct extent_state **cached_ptr,
 830                                  unsigned flags)
 831 {
 832         if (cached_ptr && !(*cached_ptr)) {
 833                 if (!flags || (state->state & flags)) {
 834                         *cached_ptr = state;
 835                         atomic_inc(&state->refs);
 836                 }
 837         }
 838 }
 839
 840 static void cache_state(struct extent_state *state,
 841                         struct extent_state **cached_ptr)
 842 {
 843         return cache_state_if_flags(state, cached_ptr,
 844                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 845 }
 846
 847 /*
 848  * set some bits on a range in the tree.  This may require allocations or
 849  * sleeping, so the gfp mask is used to indicate what is allowed.
 850  *
 851  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 852  * part of the range already has the desired bits set.  The start of the
 853  * existing range is returned in failed_start in this case.
 854  *
 855  * [start, end] is inclusive This takes the tree lock.
 856  */
 857
 858 static int __must_check
 859 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 860                  unsigned bits, unsigned exclusive_bits,
 861                  u64 *failed_start, struct extent_state **cached_state,
 862                  gfp_t mask, struct extent_changeset *changeset)
 863 {
 864         struct extent_state *state;
 865         struct extent_state *prealloc = NULL;
 866         struct rb_node *node;
 867         struct rb_node **p;
 868         struct rb_node *parent;
 869         int err = 0;
 870         u64 last_start;
 871         u64 last_end;
 872
 873         btrfs_debug_check_extent_io_range(tree, start, end);
 874
 875         bits |= EXTENT_FIRST_DELALLOC;
 876 again:
 877         if (!prealloc && gfpflags_allow_blocking(mask)) {
 878                 prealloc = alloc_extent_state(mask);
 879                 BUG_ON(!prealloc);
 880         }
 881
 882         spin_lock(&tree->lock);
 883         if (cached_state && *cached_state) {
 884                 state = *cached_state;
 885                 if (state->start <= start && state->end > start &&
 886                     extent_state_in_tree(state)) {
 887                         node = &state->rb_node;
 888                         goto hit_next;
 889                 }
 890         }
 891         /*
 892          * this search will find all the extents that end after
 893          * our range starts.
 894          */
 895         node = tree_search_for_insert(tree, start, &p, &parent);
 896         if (!node) {
 897                 prealloc = alloc_extent_state_atomic(prealloc);
 898                 BUG_ON(!prealloc);
 899                 err = insert_state(tree, prealloc, start, end,
 900                                    &p, &parent, &bits, changeset);
 901                 if (err)
 902                         extent_io_tree_panic(tree, err);
 903
 904                 cache_state(prealloc, cached_state);
 905                 prealloc = NULL;
 906                 goto out;
 907         }
 908         state = rb_entry(node, struct extent_state, rb_node);
 909 hit_next:
 910         last_start = state->start;
 911         last_end = state->end;
 912
 913         /*
 914          * | ---- desired range ---- |
 915          * | state |
 916          *
 917          * Just lock what we found and keep going
 918          */
 919         if (state->start == start && state->end <= end) {
 920                 if (state->state & exclusive_bits) {
 921                         *failed_start = state->start;
 922                         err = -EEXIST;
 923                         goto out;
 924                 }
 925
 926                 set_state_bits(tree, state, &bits, changeset);
 927                 cache_state(state, cached_state);
 928                 merge_state(tree, state);
 929                 if (last_end == (u64)-1)
 930                         goto out;
 931                 start = last_end + 1;
 932                 state = next_state(state);
 933                 if (start < end && state && state->start == start &&
 934                     !need_resched())
 935                         goto hit_next;
 936                 goto search_again;
 937         }
 938
 939         /*
 940          *     | ---- desired range ---- |
 941          * | state |
 942          *   or
 943          * | ------------- state -------------- |
 944          *
 945          * We need to split the extent we found, and may flip bits on
 946          * second half.
 947          *
 948          * If the extent we found extends past our
 949          * range, we just split and search again.  It'll get split
 950          * again the next time though.
 951          *
 952          * If the extent we found is inside our range, we set the
 953          * desired bit on it.
 954          */
 955         if (state->start < start) {
 956                 if (state->state & exclusive_bits) {
 957                         *failed_start = start;
 958                         err = -EEXIST;
 959                         goto out;
 960                 }
 961
 962                 prealloc = alloc_extent_state_atomic(prealloc);
 963                 BUG_ON(!prealloc);
 964                 err = split_state(tree, state, prealloc, start);
 965                 if (err)
 966                         extent_io_tree_panic(tree, err);
 967
 968                 prealloc = NULL;
 969                 if (err)
 970                         goto out;
 971                 if (state->end <= end) {
 972                         set_state_bits(tree, state, &bits, changeset);
 973                         cache_state(state, cached_state);
 974                         merge_state(tree, state);
 975                         if (last_end == (u64)-1)
 976                                 goto out;
 977                         start = last_end + 1;
 978                         state = next_state(state);
 979                         if (start < end && state && state->start == start &&
 980                             !need_resched())
 981                                 goto hit_next;
 982                 }
 983                 goto search_again;
 984         }
 985         /*
 986          * | ---- desired range ---- |
 987          *     | state | or               | state |
 988          *
 989          * There's a hole, we need to insert something in it and
 990          * ignore the extent we found.
 991          */
 992         if (state->start > start) {
 993                 u64 this_end;
 994                 if (end < last_start)
 995                         this_end = end;
 996                 else
 997                         this_end = last_start - 1;
 998
 999                 prealloc = alloc_extent_state_atomic(prealloc);
1000                 BUG_ON(!prealloc);
1001
1002                 /*
1003                  * Avoid to free 'prealloc' if it can be merged with
1004                  * the later extent.
1005                  */
1006                 err = insert_state(tree, prealloc, start, this_end,
1007                                    NULL, NULL, &bits, changeset);
1008                 if (err)
1009                         extent_io_tree_panic(tree, err);
1010
1011                 cache_state(prealloc, cached_state);
1012                 prealloc = NULL;
1013                 start = this_end + 1;
1014                 goto search_again;
1015         }
1016         /*
1017          * | ---- desired range ---- |
1018          *                        | state |
1019          * We need to split the extent, and set the bit
1020          * on the first half
1021          */
1022         if (state->start <= end && state->end > end) {
1023                 if (state->state & exclusive_bits) {
1024                         *failed_start = start;
1025                         err = -EEXIST;
1026                         goto out;
1027                 }
1028
1029                 prealloc = alloc_extent_state_atomic(prealloc);
1030                 BUG_ON(!prealloc);
1031                 err = split_state(tree, state, prealloc, end + 1);
1032                 if (err)
1033                         extent_io_tree_panic(tree, err);
1034
1035                 set_state_bits(tree, prealloc, &bits, changeset);
1036                 cache_state(prealloc, cached_state);
1037                 merge_state(tree, prealloc);
1038                 prealloc = NULL;
1039                 goto out;
1040         }
1041
1042         goto search_again;
1043
1044 out:
1045         spin_unlock(&tree->lock);
1046         if (prealloc)
1047                 free_extent_state(prealloc);
1048
1049         return err;
1050
1051 search_again:
1052         if (start > end)
1053                 goto out;
1054         spin_unlock(&tree->lock);
1055         if (gfpflags_allow_blocking(mask))
1056                 cond_resched();
1057         goto again;
1058 }
1059
1060 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1061                    unsigned bits, u64 * failed_start,
1062                    struct extent_state **cached_state, gfp_t mask)
1063 {
1064         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1065                                 cached_state, mask, NULL);
1066 }
1067
1068
1069 /**
1070  * convert_extent_bit - convert all bits in a given range from one bit to
1071  *                      another
1072  * @tree:       the io tree to search
1073  * @start:      the start offset in bytes
1074  * @end:        the end offset in bytes (inclusive)
1075  * @bits:       the bits to set in this range
1076  * @clear_bits: the bits to clear in this range
1077  * @cached_state:       state that we're going to cache
1078  * @mask:       the allocation mask
1079  *
1080  * This will go through and set bits for the given range.  If any states exist
1081  * already in this range they are set with the given bit and cleared of the
1082  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1083  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1084  * boundary bits like LOCK.
1085  */
1086 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1087                        unsigned bits, unsigned clear_bits,
1088                        struct extent_state **cached_state, gfp_t mask)
1089 {
1090         struct extent_state *state;
1091         struct extent_state *prealloc = NULL;
1092         struct rb_node *node;
1093         struct rb_node **p;
1094         struct rb_node *parent;
1095         int err = 0;
1096         u64 last_start;
1097         u64 last_end;
1098         bool first_iteration = true;
1099
1100         btrfs_debug_check_extent_io_range(tree, start, end);
1101
1102 again:
1103         if (!prealloc && gfpflags_allow_blocking(mask)) {
1104                 /*
1105                  * Best effort, don't worry if extent state allocation fails
1106                  * here for the first iteration. We might have a cached state
1107                  * that matches exactly the target range, in which case no
1108                  * extent state allocations are needed. We'll only know this
1109                  * after locking the tree.
1110                  */
1111                 prealloc = alloc_extent_state(mask);
1112                 if (!prealloc && !first_iteration)
1113                         return -ENOMEM;
1114         }
1115
1116         spin_lock(&tree->lock);
1117         if (cached_state && *cached_state) {
1118                 state = *cached_state;
1119                 if (state->start <= start && state->end > start &&
1120                     extent_state_in_tree(state)) {
1121                         node = &state->rb_node;
1122                         goto hit_next;
1123                 }
1124         }
1125
1126         /*
1127          * this search will find all the extents that end after
1128          * our range starts.
1129          */
1130         node = tree_search_for_insert(tree, start, &p, &parent);
1131         if (!node) {
1132                 prealloc = alloc_extent_state_atomic(prealloc);
1133                 if (!prealloc) {
1134                         err = -ENOMEM;
1135                         goto out;
1136                 }
1137                 err = insert_state(tree, prealloc, start, end,
1138                                    &p, &parent, &bits, NULL);
1139                 if (err)
1140                         extent_io_tree_panic(tree, err);
1141                 cache_state(prealloc, cached_state);
1142                 prealloc = NULL;
1143                 goto out;
1144         }
1145         state = rb_entry(node, struct extent_state, rb_node);
1146 hit_next:
1147         last_start = state->start;
1148         last_end = state->end;
1149
1150         /*
1151          * | ---- desired range ---- |
1152          * | state |
1153          *
1154          * Just lock what we found and keep going
1155          */
1156         if (state->start == start && state->end <= end) {
1157                 set_state_bits(tree, state, &bits, NULL);
1158                 cache_state(state, cached_state);
1159                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1160                 if (last_end == (u64)-1)
1161                         goto out;
1162                 start = last_end + 1;
1163                 if (start < end && state && state->start == start &&
1164                     !need_resched())
1165                         goto hit_next;
1166                 goto search_again;
1167         }
1168
1169         /*
1170          *     | ---- desired range ---- |
1171          * | state |
1172          *   or
1173          * | ------------- state -------------- |
1174          *
1175          * We need to split the extent we found, and may flip bits on
1176          * second half.
1177          *
1178          * If the extent we found extends past our
1179          * range, we just split and search again.  It'll get split
1180          * again the next time though.
1181          *
1182          * If the extent we found is inside our range, we set the
1183          * desired bit on it.
1184          */
1185         if (state->start < start) {
1186                 prealloc = alloc_extent_state_atomic(prealloc);
1187                 if (!prealloc) {
1188                         err = -ENOMEM;
1189                         goto out;
1190                 }
1191                 err = split_state(tree, state, prealloc, start);
1192                 if (err)
1193                         extent_io_tree_panic(tree, err);
1194                 prealloc = NULL;
1195                 if (err)
1196                         goto out;
1197                 if (state->end <= end) {
1198                         set_state_bits(tree, state, &bits, NULL);
1199                         cache_state(state, cached_state);
1200                         state = clear_state_bit(tree, state, &clear_bits, 0,
1201                                                 NULL);
1202                         if (last_end == (u64)-1)
1203                                 goto out;
1204                         start = last_end + 1;
1205                         if (start < end && state && state->start == start &&
1206                             !need_resched())
1207                                 goto hit_next;
1208                 }
1209                 goto search_again;
1210         }
1211         /*
1212          * | ---- desired range ---- |
1213          *     | state | or               | state |
1214          *
1215          * There's a hole, we need to insert something in it and
1216          * ignore the extent we found.
1217          */
1218         if (state->start > start) {
1219                 u64 this_end;
1220                 if (end < last_start)
1221                         this_end = end;
1222                 else
1223                         this_end = last_start - 1;
1224
1225                 prealloc = alloc_extent_state_atomic(prealloc);
1226                 if (!prealloc) {
1227                         err = -ENOMEM;
1228                         goto out;
1229                 }
1230
1231                 /*
1232                  * Avoid to free 'prealloc' if it can be merged with
1233                  * the later extent.
1234                  */
1235                 err = insert_state(tree, prealloc, start, this_end,
1236                                    NULL, NULL, &bits, NULL);
1237                 if (err)
1238                         extent_io_tree_panic(tree, err);
1239                 cache_state(prealloc, cached_state);
1240                 prealloc = NULL;
1241                 start = this_end + 1;
1242                 goto search_again;
1243         }
1244         /*
1245          * | ---- desired range ---- |
1246          *                        | state |
1247          * We need to split the extent, and set the bit
1248          * on the first half
1249          */
1250         if (state->start <= end && state->end > end) {
1251                 prealloc = alloc_extent_state_atomic(prealloc);
1252                 if (!prealloc) {
1253                         err = -ENOMEM;
1254                         goto out;
1255                 }
1256
1257                 err = split_state(tree, state, prealloc, end + 1);
1258                 if (err)
1259                         extent_io_tree_panic(tree, err);
1260
1261                 set_state_bits(tree, prealloc, &bits, NULL);
1262                 cache_state(prealloc, cached_state);
1263                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1264                 prealloc = NULL;
1265                 goto out;
1266         }
1267
1268         goto search_again;
1269
1270 out:
1271         spin_unlock(&tree->lock);
1272         if (prealloc)
1273                 free_extent_state(prealloc);
1274
1275         return err;
1276
1277 search_again:
1278         if (start > end)
1279                 goto out;
1280         spin_unlock(&tree->lock);
1281         if (gfpflags_allow_blocking(mask))
1282                 cond_resched();
1283         first_iteration = false;
1284         goto again;
1285 }
1286
1287 /* wrappers around set/clear extent bit */
1288 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1289                      gfp_t mask)
1290 {
1291         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1292                               NULL, mask);
1293 }
1294
1295 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1296                     unsigned bits, gfp_t mask)
1297 {
1298         return set_extent_bit(tree, start, end, bits, NULL,
1299                               NULL, mask);
1300 }
1301
1302 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1303                            unsigned bits, gfp_t mask,
1304                            struct extent_changeset *changeset)
1305 {
1306         /*
1307          * We don't support EXTENT_LOCKED yet, as current changeset will
1308          * record any bits changed, so for EXTENT_LOCKED case, it will
1309          * either fail with -EEXIST or changeset will record the whole
1310          * range.
1311          */
1312         BUG_ON(bits & EXTENT_LOCKED);
1313
1314         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, mask,
1315                                 changeset);
1316 }
1317
1318 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1319                      unsigned bits, int wake, int delete,
1320                      struct extent_state **cached, gfp_t mask)
1321 {
1322         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1323                                   cached, mask, NULL);
1324 }
1325
1326 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1327                       unsigned bits, gfp_t mask)
1328 {
1329         int wake = 0;
1330
1331         if (bits & EXTENT_LOCKED)
1332                 wake = 1;
1333
1334         return clear_extent_bit(tree, start, end, bits, wake, 0, NULL, mask);
1335 }
1336
1337 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1338                              unsigned bits, gfp_t mask,
1339                              struct extent_changeset *changeset)
1340 {
1341         /*
1342          * Don't support EXTENT_LOCKED case, same reason as
1343          * set_record_extent_bits().
1344          */
1345         BUG_ON(bits & EXTENT_LOCKED);
1346
1347         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask,
1348                                   changeset);
1349 }
1350
1351 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1352                         struct extent_state **cached_state, gfp_t mask)
1353 {
1354         return set_extent_bit(tree, start, end,
1355                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1356                               NULL, cached_state, mask);
1357 }
1358
1359 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1360                       struct extent_state **cached_state, gfp_t mask)
1361 {
1362         return set_extent_bit(tree, start, end,
1363                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1364                               NULL, cached_state, mask);
1365 }
1366
1367 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1368                        gfp_t mask)
1369 {
1370         return clear_extent_bit(tree, start, end,
1371                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1372                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1373 }
1374
1375 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1376                      gfp_t mask)
1377 {
1378         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1379                               NULL, mask);
1380 }
1381
1382 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1383                         struct extent_state **cached_state, gfp_t mask)
1384 {
1385         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1386                               cached_state, mask);
1387 }
1388
1389 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1390                           struct extent_state **cached_state, gfp_t mask)
1391 {
1392         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1393                                 cached_state, mask);
1394 }
1395
1396 /*
1397  * either insert or lock state struct between start and end use mask to tell
1398  * us if waiting is desired.
1399  */
1400 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1401                      unsigned bits, struct extent_state **cached_state)
1402 {
1403         int err;
1404         u64 failed_start;
1405
1406         while (1) {
1407                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1408                                        EXTENT_LOCKED, &failed_start,
1409                                        cached_state, GFP_NOFS, NULL);
1410                 if (err == -EEXIST) {
1411                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1412                         start = failed_start;
1413                 } else
1414                         break;
1415                 WARN_ON(start > end);
1416         }
1417         return err;
1418 }
1419
1420 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1421 {
1422         return lock_extent_bits(tree, start, end, 0, NULL);
1423 }
1424
1425 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1426 {
1427         int err;
1428         u64 failed_start;
1429
1430         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1431                                &failed_start, NULL, GFP_NOFS, NULL);
1432         if (err == -EEXIST) {
1433                 if (failed_start > start)
1434                         clear_extent_bit(tree, start, failed_start - 1,
1435                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1436                 return 0;
1437         }
1438         return 1;
1439 }
1440
1441 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1442                          struct extent_state **cached, gfp_t mask)
1443 {
1444         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1445                                 mask);
1446 }
1447
1448 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1449 {
1450         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1451                                 GFP_NOFS);
1452 }
1453
1454 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1455 {
1456         unsigned long index = start >> PAGE_CACHE_SHIFT;
1457         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1458         struct page *page;
1459
1460         while (index <= end_index) {
1461                 page = find_get_page(inode->i_mapping, index);
1462                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1463                 clear_page_dirty_for_io(page);
1464                 page_cache_release(page);
1465                 index++;
1466         }
1467         return 0;
1468 }
1469
1470 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1471 {
1472         unsigned long index = start >> PAGE_CACHE_SHIFT;
1473         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1474         struct page *page;
1475
1476         while (index <= end_index) {
1477                 page = find_get_page(inode->i_mapping, index);
1478                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1479                 __set_page_dirty_nobuffers(page);
1480                 account_page_redirty(page);
1481                 page_cache_release(page);
1482                 index++;
1483         }
1484         return 0;
1485 }
1486
1487 /*
1488  * helper function to set both pages and extents in the tree writeback
1489  */
1490 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1491 {
1492         unsigned long index = start >> PAGE_CACHE_SHIFT;
1493         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1494         struct page *page;
1495
1496         while (index <= end_index) {
1497                 page = find_get_page(tree->mapping, index);
1498                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1499                 set_page_writeback(page);
1500                 page_cache_release(page);
1501                 index++;
1502         }
1503         return 0;
1504 }
1505
1506 /* find the first state struct with 'bits' set after 'start', and
1507  * return it.  tree->lock must be held.  NULL will returned if
1508  * nothing was found after 'start'
1509  */
1510 static struct extent_state *
1511 find_first_extent_bit_state(struct extent_io_tree *tree,
1512                             u64 start, unsigned bits)
1513 {
1514         struct rb_node *node;
1515         struct extent_state *state;
1516
1517         /*
1518          * this search will find all the extents that end after
1519          * our range starts.
1520          */
1521         node = tree_search(tree, start);
1522         if (!node)
1523                 goto out;
1524
1525         while (1) {
1526                 state = rb_entry(node, struct extent_state, rb_node);
1527                 if (state->end >= start && (state->state & bits))
1528                         return state;
1529
1530                 node = rb_next(node);
1531                 if (!node)
1532                         break;
1533         }
1534 out:
1535         return NULL;
1536 }
1537
1538 /*
1539  * find the first offset in the io tree with 'bits' set. zero is
1540  * returned if we find something, and *start_ret and *end_ret are
1541  * set to reflect the state struct that was found.
1542  *
1543  * If nothing was found, 1 is returned. If found something, return 0.
1544  */
1545 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1546                           u64 *start_ret, u64 *end_ret, unsigned bits,
1547                           struct extent_state **cached_state)
1548 {
1549         struct extent_state *state;
1550         struct rb_node *n;
1551         int ret = 1;
1552
1553         spin_lock(&tree->lock);
1554         if (cached_state && *cached_state) {
1555                 state = *cached_state;
1556                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1557                         n = rb_next(&state->rb_node);
1558                         while (n) {
1559                                 state = rb_entry(n, struct extent_state,
1560                                                  rb_node);
1561                                 if (state->state & bits)
1562                                         goto got_it;
1563                                 n = rb_next(n);
1564                         }
1565                         free_extent_state(*cached_state);
1566                         *cached_state = NULL;
1567                         goto out;
1568                 }
1569                 free_extent_state(*cached_state);
1570                 *cached_state = NULL;
1571         }
1572
1573         state = find_first_extent_bit_state(tree, start, bits);
1574 got_it:
1575         if (state) {
1576                 cache_state_if_flags(state, cached_state, 0);
1577                 *start_ret = state->start;
1578                 *end_ret = state->end;
1579                 ret = 0;
1580         }
1581 out:
1582         spin_unlock(&tree->lock);
1583         return ret;
1584 }
1585
1586 /*
1587  * find a contiguous range of bytes in the file marked as delalloc, not
1588  * more than 'max_bytes'.  start and end are used to return the range,
1589  *
1590  * 1 is returned if we find something, 0 if nothing was in the tree
1591  */
1592 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1593                                         u64 *start, u64 *end, u64 max_bytes,
1594                                         struct extent_state **cached_state)
1595 {
1596         struct rb_node *node;
1597         struct extent_state *state;
1598         u64 cur_start = *start;
1599         u64 found = 0;
1600         u64 total_bytes = 0;
1601
1602         spin_lock(&tree->lock);
1603
1604         /*
1605          * this search will find all the extents that end after
1606          * our range starts.
1607          */
1608         node = tree_search(tree, cur_start);
1609         if (!node) {
1610                 if (!found)
1611                         *end = (u64)-1;
1612                 goto out;
1613         }
1614
1615         while (1) {
1616                 state = rb_entry(node, struct extent_state, rb_node);
1617                 if (found && (state->start != cur_start ||
1618                               (state->state & EXTENT_BOUNDARY))) {
1619                         goto out;
1620                 }
1621                 if (!(state->state & EXTENT_DELALLOC)) {
1622                         if (!found)
1623                                 *end = state->end;
1624                         goto out;
1625                 }
1626                 if (!found) {
1627                         *start = state->start;
1628                         *cached_state = state;
1629                         atomic_inc(&state->refs);
1630                 }
1631                 found++;
1632                 *end = state->end;
1633                 cur_start = state->end + 1;
1634                 node = rb_next(node);
1635                 total_bytes += state->end - state->start + 1;
1636                 if (total_bytes >= max_bytes)
1637                         break;
1638                 if (!node)
1639                         break;
1640         }
1641 out:
1642         spin_unlock(&tree->lock);
1643         return found;
1644 }
1645
1646 static noinline void __unlock_for_delalloc(struct inode *inode,
1647                                            struct page *locked_page,
1648                                            u64 start, u64 end)
1649 {
1650         int ret;
1651         struct page *pages[16];
1652         unsigned long index = start >> PAGE_CACHE_SHIFT;
1653         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1654         unsigned long nr_pages = end_index - index + 1;
1655         int i;
1656
1657         if (index == locked_page->index && end_index == index)
1658                 return;
1659
1660         while (nr_pages > 0) {
1661                 ret = find_get_pages_contig(inode->i_mapping, index,
1662                                      min_t(unsigned long, nr_pages,
1663                                      ARRAY_SIZE(pages)), pages);
1664                 for (i = 0; i < ret; i++) {
1665                         if (pages[i] != locked_page)
1666                                 unlock_page(pages[i]);
1667                         page_cache_release(pages[i]);
1668                 }
1669                 nr_pages -= ret;
1670                 index += ret;
1671                 cond_resched();
1672         }
1673 }
1674
1675 static noinline int lock_delalloc_pages(struct inode *inode,
1676                                         struct page *locked_page,
1677                                         u64 delalloc_start,
1678                                         u64 delalloc_end)
1679 {
1680         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1681         unsigned long start_index = index;
1682         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1683         unsigned long pages_locked = 0;
1684         struct page *pages[16];
1685         unsigned long nrpages;
1686         int ret;
1687         int i;
1688
1689         /* the caller is responsible for locking the start index */
1690         if (index == locked_page->index && index == end_index)
1691                 return 0;
1692
1693         /* skip the page at the start index */
1694         nrpages = end_index - index + 1;
1695         while (nrpages > 0) {
1696                 ret = find_get_pages_contig(inode->i_mapping, index,
1697                                      min_t(unsigned long,
1698                                      nrpages, ARRAY_SIZE(pages)), pages);
1699                 if (ret == 0) {
1700                         ret = -EAGAIN;
1701                         goto done;
1702                 }
1703                 /* now we have an array of pages, lock them all */
1704                 for (i = 0; i < ret; i++) {
1705                         /*
1706                          * the caller is taking responsibility for
1707                          * locked_page
1708                          */
1709                         if (pages[i] != locked_page) {
1710                                 lock_page(pages[i]);
1711                                 if (!PageDirty(pages[i]) ||
1712                                     pages[i]->mapping != inode->i_mapping) {
1713                                         ret = -EAGAIN;
1714                                         unlock_page(pages[i]);
1715                                         page_cache_release(pages[i]);
1716                                         goto done;
1717                                 }
1718                         }
1719                         page_cache_release(pages[i]);
1720                         pages_locked++;
1721                 }
1722                 nrpages -= ret;
1723                 index += ret;
1724                 cond_resched();
1725         }
1726         ret = 0;
1727 done:
1728         if (ret && pages_locked) {
1729                 __unlock_for_delalloc(inode, locked_page,
1730                               delalloc_start,
1731                               ((u64)(start_index + pages_locked - 1)) <<
1732                               PAGE_CACHE_SHIFT);
1733         }
1734         return ret;
1735 }
1736
1737 /*
1738  * find a contiguous range of bytes in the file marked as delalloc, not
1739  * more than 'max_bytes'.  start and end are used to return the range,
1740  *
1741  * 1 is returned if we find something, 0 if nothing was in the tree
1742  */
1743 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1744                                     struct extent_io_tree *tree,
1745                                     struct page *locked_page, u64 *start,
1746                                     u64 *end, u64 max_bytes)
1747 {
1748         u64 delalloc_start;
1749         u64 delalloc_end;
1750         u64 found;
1751         struct extent_state *cached_state = NULL;
1752         int ret;
1753         int loops = 0;
1754
1755 again:
1756         /* step one, find a bunch of delalloc bytes starting at start */
1757         delalloc_start = *start;
1758         delalloc_end = 0;
1759         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1760                                     max_bytes, &cached_state);
1761         if (!found || delalloc_end <= *start) {
1762                 *start = delalloc_start;
1763                 *end = delalloc_end;
1764                 free_extent_state(cached_state);
1765                 return 0;
1766         }
1767
1768         /*
1769          * start comes from the offset of locked_page.  We have to lock
1770          * pages in order, so we can't process delalloc bytes before
1771          * locked_page
1772          */
1773         if (delalloc_start < *start)
1774                 delalloc_start = *start;
1775
1776         /*
1777          * make sure to limit the number of pages we try to lock down
1778          */
1779         if (delalloc_end + 1 - delalloc_start > max_bytes)
1780                 delalloc_end = delalloc_start + max_bytes - 1;
1781
1782         /* step two, lock all the pages after the page that has start */
1783         ret = lock_delalloc_pages(inode, locked_page,
1784                                   delalloc_start, delalloc_end);
1785         if (ret == -EAGAIN) {
1786                 /* some of the pages are gone, lets avoid looping by
1787                  * shortening the size of the delalloc range we're searching
1788                  */
1789                 free_extent_state(cached_state);
1790                 cached_state = NULL;
1791                 if (!loops) {
1792                         max_bytes = PAGE_CACHE_SIZE;
1793                         loops = 1;
1794                         goto again;
1795                 } else {
1796                         found = 0;
1797                         goto out_failed;
1798                 }
1799         }
1800         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1801
1802         /* step three, lock the state bits for the whole range */
1803         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1804
1805         /* then test to make sure it is all still delalloc */
1806         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1807                              EXTENT_DELALLOC, 1, cached_state);
1808         if (!ret) {
1809                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1810                                      &cached_state, GFP_NOFS);
1811                 __unlock_for_delalloc(inode, locked_page,
1812                               delalloc_start, delalloc_end);
1813                 cond_resched();
1814                 goto again;
1815         }
1816         free_extent_state(cached_state);
1817         *start = delalloc_start;
1818         *end = delalloc_end;
1819 out_failed:
1820         return found;
1821 }
1822
1823 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1824                                  struct page *locked_page,
1825                                  unsigned clear_bits,
1826                                  unsigned long page_ops)
1827 {
1828         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1829         int ret;
1830         struct page *pages[16];
1831         unsigned long index = start >> PAGE_CACHE_SHIFT;
1832         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1833         unsigned long nr_pages = end_index - index + 1;
1834         int i;
1835
1836         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1837         if (page_ops == 0)
1838                 return 0;
1839
1840         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1841                 mapping_set_error(inode->i_mapping, -EIO);
1842
1843         while (nr_pages > 0) {
1844                 ret = find_get_pages_contig(inode->i_mapping, index,
1845                                      min_t(unsigned long,
1846                                      nr_pages, ARRAY_SIZE(pages)), pages);
1847                 for (i = 0; i < ret; i++) {
1848
1849                         if (page_ops & PAGE_SET_PRIVATE2)
1850                                 SetPagePrivate2(pages[i]);
1851
1852                         if (pages[i] == locked_page) {
1853                                 page_cache_release(pages[i]);
1854                                 continue;
1855                         }
1856                         if (page_ops & PAGE_CLEAR_DIRTY)
1857                                 clear_page_dirty_for_io(pages[i]);
1858                         if (page_ops & PAGE_SET_WRITEBACK)
1859                                 set_page_writeback(pages[i]);
1860                         if (page_ops & PAGE_SET_ERROR)
1861                                 SetPageError(pages[i]);
1862                         if (page_ops & PAGE_END_WRITEBACK)
1863                                 end_page_writeback(pages[i]);
1864                         if (page_ops & PAGE_UNLOCK)
1865                                 unlock_page(pages[i]);
1866                         page_cache_release(pages[i]);
1867                 }
1868                 nr_pages -= ret;
1869                 index += ret;
1870                 cond_resched();
1871         }
1872         return 0;
1873 }
1874
1875 /*
1876  * count the number of bytes in the tree that have a given bit(s)
1877  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1878  * cached.  The total number found is returned.
1879  */
1880 u64 count_range_bits(struct extent_io_tree *tree,
1881                      u64 *start, u64 search_end, u64 max_bytes,
1882                      unsigned bits, int contig)
1883 {
1884         struct rb_node *node;
1885         struct extent_state *state;
1886         u64 cur_start = *start;
1887         u64 total_bytes = 0;
1888         u64 last = 0;
1889         int found = 0;
1890
1891         if (WARN_ON(search_end <= cur_start))
1892                 return 0;
1893
1894         spin_lock(&tree->lock);
1895         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1896                 total_bytes = tree->dirty_bytes;
1897                 goto out;
1898         }
1899         /*
1900          * this search will find all the extents that end after
1901          * our range starts.
1902          */
1903         node = tree_search(tree, cur_start);
1904         if (!node)
1905                 goto out;
1906
1907         while (1) {
1908                 state = rb_entry(node, struct extent_state, rb_node);
1909                 if (state->start > search_end)
1910                         break;
1911                 if (contig && found && state->start > last + 1)
1912                         break;
1913                 if (state->end >= cur_start && (state->state & bits) == bits) {
1914                         total_bytes += min(search_end, state->end) + 1 -
1915                                        max(cur_start, state->start);
1916                         if (total_bytes >= max_bytes)
1917                                 break;
1918                         if (!found) {
1919                                 *start = max(cur_start, state->start);
1920                                 found = 1;
1921                         }
1922                         last = state->end;
1923                 } else if (contig && found) {
1924                         break;
1925                 }
1926                 node = rb_next(node);
1927                 if (!node)
1928                         break;
1929         }
1930 out:
1931         spin_unlock(&tree->lock);
1932         return total_bytes;
1933 }
1934
1935 /*
1936  * set the private field for a given byte offset in the tree.  If there isn't
1937  * an extent_state there already, this does nothing.
1938  */
1939 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1940 {
1941         struct rb_node *node;
1942         struct extent_state *state;
1943         int ret = 0;
1944
1945         spin_lock(&tree->lock);
1946         /*
1947          * this search will find all the extents that end after
1948          * our range starts.
1949          */
1950         node = tree_search(tree, start);
1951         if (!node) {
1952                 ret = -ENOENT;
1953                 goto out;
1954         }
1955         state = rb_entry(node, struct extent_state, rb_node);
1956         if (state->start != start) {
1957                 ret = -ENOENT;
1958                 goto out;
1959         }
1960         state->private = private;
1961 out:
1962         spin_unlock(&tree->lock);
1963         return ret;
1964 }
1965
1966 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1967 {
1968         struct rb_node *node;
1969         struct extent_state *state;
1970         int ret = 0;
1971
1972         spin_lock(&tree->lock);
1973         /*
1974          * this search will find all the extents that end after
1975          * our range starts.
1976          */
1977         node = tree_search(tree, start);
1978         if (!node) {
1979                 ret = -ENOENT;
1980                 goto out;
1981         }
1982         state = rb_entry(node, struct extent_state, rb_node);
1983         if (state->start != start) {
1984                 ret = -ENOENT;
1985                 goto out;
1986         }
1987         *private = state->private;
1988 out:
1989         spin_unlock(&tree->lock);
1990         return ret;
1991 }
1992
1993 /*
1994  * searches a range in the state tree for a given mask.
1995  * If 'filled' == 1, this returns 1 only if every extent in the tree
1996  * has the bits set.  Otherwise, 1 is returned if any bit in the
1997  * range is found set.
1998  */
1999 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
2000                    unsigned bits, int filled, struct extent_state *cached)
2001 {
2002         struct extent_state *state = NULL;
2003         struct rb_node *node;
2004         int bitset = 0;
2005
2006         spin_lock(&tree->lock);
2007         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
2008             cached->end > start)
2009                 node = &cached->rb_node;
2010         else
2011                 node = tree_search(tree, start);
2012         while (node && start <= end) {
2013                 state = rb_entry(node, struct extent_state, rb_node);
2014
2015                 if (filled && state->start > start) {
2016                         bitset = 0;
2017                         break;
2018                 }
2019
2020                 if (state->start > end)
2021                         break;
2022
2023                 if (state->state & bits) {
2024                         bitset = 1;
2025                         if (!filled)
2026                                 break;
2027                 } else if (filled) {
2028                         bitset = 0;
2029                         break;
2030                 }
2031
2032                 if (state->end == (u64)-1)
2033                         break;
2034
2035                 start = state->end + 1;
2036                 if (start > end)
2037                         break;
2038                 node = rb_next(node);
2039                 if (!node) {
2040                         if (filled)
2041                                 bitset = 0;
2042                         break;
2043                 }
2044         }
2045         spin_unlock(&tree->lock);
2046         return bitset;
2047 }
2048
2049 /*
2050  * helper function to set a given page up to date if all the
2051  * extents in the tree for that page are up to date
2052  */
2053 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
2054 {
2055         u64 start = page_offset(page);
2056         u64 end = start + PAGE_CACHE_SIZE - 1;
2057         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
2058                 SetPageUptodate(page);
2059 }
2060
2061 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
2062 {
2063         int ret;
2064         int err = 0;
2065         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2066
2067         set_state_private(failure_tree, rec->start, 0);
2068         ret = clear_extent_bits(failure_tree, rec->start,
2069                                 rec->start + rec->len - 1,
2070                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2071         if (ret)
2072                 err = ret;
2073
2074         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
2075                                 rec->start + rec->len - 1,
2076                                 EXTENT_DAMAGED, GFP_NOFS);
2077         if (ret && !err)
2078                 err = ret;
2079
2080         kfree(rec);
2081         return err;
2082 }
2083
2084 /*
2085  * this bypasses the standard btrfs submit functions deliberately, as
2086  * the standard behavior is to write all copies in a raid setup. here we only
2087  * want to write the one bad copy. so we do the mapping for ourselves and issue
2088  * submit_bio directly.
2089  * to avoid any synchronization issues, wait for the data after writing, which
2090  * actually prevents the read that triggered the error from finishing.
2091  * currently, there can be no more than two copies of every data bit. thus,
2092  * exactly one rewrite is required.
2093  */
2094 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2095                       struct page *page, unsigned int pg_offset, int mirror_num)
2096 {
2097         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2098         struct bio *bio;
2099         struct btrfs_device *dev;
2100         u64 map_length = 0;
2101         u64 sector;
2102         struct btrfs_bio *bbio = NULL;
2103         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2104         int ret;
2105
2106         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2107         BUG_ON(!mirror_num);
2108
2109         /* we can't repair anything in raid56 yet */
2110         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2111                 return 0;
2112
2113         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2114         if (!bio)
2115                 return -EIO;
2116         bio->bi_iter.bi_size = 0;
2117         map_length = length;
2118
2119         ret = btrfs_map_block(fs_info, WRITE, logical,
2120                               &map_length, &bbio, mirror_num);
2121         if (ret) {
2122                 bio_put(bio);
2123                 return -EIO;
2124         }
2125         BUG_ON(mirror_num != bbio->mirror_num);
2126         sector = bbio->stripes[mirror_num-1].physical >> 9;
2127         bio->bi_iter.bi_sector = sector;
2128         dev = bbio->stripes[mirror_num-1].dev;
2129         btrfs_put_bbio(bbio);
2130         if (!dev || !dev->bdev || !dev->writeable) {
2131                 bio_put(bio);
2132                 return -EIO;
2133         }
2134         bio->bi_bdev = dev->bdev;
2135         bio_add_page(bio, page, length, pg_offset);
2136
2137         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio)) {
2138                 /* try to remap that extent elsewhere? */
2139                 bio_put(bio);
2140                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2141                 return -EIO;
2142         }
2143
2144         btrfs_info_rl_in_rcu(fs_info,
2145                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2146                                   btrfs_ino(inode), start,
2147                                   rcu_str_deref(dev->name), sector);
2148         bio_put(bio);
2149         return 0;
2150 }
2151
2152 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2153                          int mirror_num)
2154 {
2155         u64 start = eb->start;
2156         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2157         int ret = 0;
2158
2159         if (root->fs_info->sb->s_flags & MS_RDONLY)
2160                 return -EROFS;
2161
2162         for (i = 0; i < num_pages; i++) {
2163                 struct page *p = eb->pages[i];
2164
2165                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2166                                         PAGE_CACHE_SIZE, start, p,
2167                                         start - page_offset(p), mirror_num);
2168                 if (ret)
2169                         break;
2170                 start += PAGE_CACHE_SIZE;
2171         }
2172
2173         return ret;
2174 }
2175
2176 /*
2177  * each time an IO finishes, we do a fast check in the IO failure tree
2178  * to see if we need to process or clean up an io_failure_record
2179  */
2180 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2181                      unsigned int pg_offset)
2182 {
2183         u64 private;
2184         u64 private_failure;
2185         struct io_failure_record *failrec;
2186         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2187         struct extent_state *state;
2188         int num_copies;
2189         int ret;
2190
2191         private = 0;
2192         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2193                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2194         if (!ret)
2195                 return 0;
2196
2197         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2198                                 &private_failure);
2199         if (ret)
2200                 return 0;
2201
2202         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2203         BUG_ON(!failrec->this_mirror);
2204
2205         if (failrec->in_validation) {
2206                 /* there was no real error, just free the record */
2207                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2208                          failrec->start);
2209                 goto out;
2210         }
2211         if (fs_info->sb->s_flags & MS_RDONLY)
2212                 goto out;
2213
2214         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2215         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2216                                             failrec->start,
2217                                             EXTENT_LOCKED);
2218         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2219
2220         if (state && state->start <= failrec->start &&
2221             state->end >= failrec->start + failrec->len - 1) {
2222                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2223                                               failrec->len);
2224                 if (num_copies > 1)  {
2225                         repair_io_failure(inode, start, failrec->len,
2226                                           failrec->logical, page,
2227                                           pg_offset, failrec->failed_mirror);
2228                 }
2229         }
2230
2231 out:
2232         free_io_failure(inode, failrec);
2233
2234         return 0;
2235 }
2236
2237 /*
2238  * Can be called when
2239  * - hold extent lock
2240  * - under ordered extent
2241  * - the inode is freeing
2242  */
2243 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2244 {
2245         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2246         struct io_failure_record *failrec;
2247         struct extent_state *state, *next;
2248
2249         if (RB_EMPTY_ROOT(&failure_tree->state))
2250                 return;
2251
2252         spin_lock(&failure_tree->lock);
2253         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2254         while (state) {
2255                 if (state->start > end)
2256                         break;
2257
2258                 ASSERT(state->end <= end);
2259
2260                 next = next_state(state);
2261
2262                 failrec = (struct io_failure_record *)(unsigned long)state->private;
2263                 free_extent_state(state);
2264                 kfree(failrec);
2265
2266                 state = next;
2267         }
2268         spin_unlock(&failure_tree->lock);
2269 }
2270
2271 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2272                                 struct io_failure_record **failrec_ret)
2273 {
2274         struct io_failure_record *failrec;
2275         u64 private;
2276         struct extent_map *em;
2277         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2278         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2279         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2280         int ret;
2281         u64 logical;
2282
2283         ret = get_state_private(failure_tree, start, &private);
2284         if (ret) {
2285                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2286                 if (!failrec)
2287                         return -ENOMEM;
2288
2289                 failrec->start = start;
2290                 failrec->len = end - start + 1;
2291                 failrec->this_mirror = 0;
2292                 failrec->bio_flags = 0;
2293                 failrec->in_validation = 0;
2294
2295                 read_lock(&em_tree->lock);
2296                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2297                 if (!em) {
2298                         read_unlock(&em_tree->lock);
2299                         kfree(failrec);
2300                         return -EIO;
2301                 }
2302
2303                 if (em->start > start || em->start + em->len <= start) {
2304                         free_extent_map(em);
2305                         em = NULL;
2306                 }
2307                 read_unlock(&em_tree->lock);
2308                 if (!em) {
2309                         kfree(failrec);
2310                         return -EIO;
2311                 }
2312
2313                 logical = start - em->start;
2314                 logical = em->block_start + logical;
2315                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2316                         logical = em->block_start;
2317                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2318                         extent_set_compress_type(&failrec->bio_flags,
2319                                                  em->compress_type);
2320                 }
2321
2322                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2323                          logical, start, failrec->len);
2324
2325                 failrec->logical = logical;
2326                 free_extent_map(em);
2327
2328                 /* set the bits in the private failure tree */
2329                 ret = set_extent_bits(failure_tree, start, end,
2330                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2331                 if (ret >= 0)
2332                         ret = set_state_private(failure_tree, start,
2333                                                 (u64)(unsigned long)failrec);
2334                 /* set the bits in the inode's tree */
2335                 if (ret >= 0)
2336                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2337                                                 GFP_NOFS);
2338                 if (ret < 0) {
2339                         kfree(failrec);
2340                         return ret;
2341                 }
2342         } else {
2343                 failrec = (struct io_failure_record *)(unsigned long)private;
2344                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2345                          failrec->logical, failrec->start, failrec->len,
2346                          failrec->in_validation);
2347                 /*
2348                  * when data can be on disk more than twice, add to failrec here
2349                  * (e.g. with a list for failed_mirror) to make
2350                  * clean_io_failure() clean all those errors at once.
2351                  */
2352         }
2353
2354         *failrec_ret = failrec;
2355
2356         return 0;
2357 }
2358
2359 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2360                            struct io_failure_record *failrec, int failed_mirror)
2361 {
2362         int num_copies;
2363
2364         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2365                                       failrec->logical, failrec->len);
2366         if (num_copies == 1) {
2367                 /*
2368                  * we only have a single copy of the data, so don't bother with
2369                  * all the retry and error correction code that follows. no
2370                  * matter what the error is, it is very likely to persist.
2371                  */
2372                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2373                          num_copies, failrec->this_mirror, failed_mirror);
2374                 return 0;
2375         }
2376
2377         /*
2378          * there are two premises:
2379          *      a) deliver good data to the caller
2380          *      b) correct the bad sectors on disk
2381          */
2382         if (failed_bio->bi_vcnt > 1) {
2383                 /*
2384                  * to fulfill b), we need to know the exact failing sectors, as
2385                  * we don't want to rewrite any more than the failed ones. thus,
2386                  * we need separate read requests for the failed bio
2387                  *
2388                  * if the following BUG_ON triggers, our validation request got
2389                  * merged. we need separate requests for our algorithm to work.
2390                  */
2391                 BUG_ON(failrec->in_validation);
2392                 failrec->in_validation = 1;
2393                 failrec->this_mirror = failed_mirror;
2394         } else {
2395                 /*
2396                  * we're ready to fulfill a) and b) alongside. get a good copy
2397                  * of the failed sector and if we succeed, we have setup
2398                  * everything for repair_io_failure to do the rest for us.
2399                  */
2400                 if (failrec->in_validation) {
2401                         BUG_ON(failrec->this_mirror != failed_mirror);
2402                         failrec->in_validation = 0;
2403                         failrec->this_mirror = 0;
2404                 }
2405                 failrec->failed_mirror = failed_mirror;
2406                 failrec->this_mirror++;
2407                 if (failrec->this_mirror == failed_mirror)
2408                         failrec->this_mirror++;
2409         }
2410
2411         if (failrec->this_mirror > num_copies) {
2412                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2413                          num_copies, failrec->this_mirror, failed_mirror);
2414                 return 0;
2415         }
2416
2417         return 1;
2418 }
2419
2420
2421 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2422                                     struct io_failure_record *failrec,
2423                                     struct page *page, int pg_offset, int icsum,
2424                                     bio_end_io_t *endio_func, void *data)
2425 {
2426         struct bio *bio;
2427         struct btrfs_io_bio *btrfs_failed_bio;
2428         struct btrfs_io_bio *btrfs_bio;
2429
2430         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2431         if (!bio)
2432                 return NULL;
2433
2434         bio->bi_end_io = endio_func;
2435         bio->bi_iter.bi_sector = failrec->logical >> 9;
2436         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2437         bio->bi_iter.bi_size = 0;
2438         bio->bi_private = data;
2439
2440         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2441         if (btrfs_failed_bio->csum) {
2442                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2443                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2444
2445                 btrfs_bio = btrfs_io_bio(bio);
2446                 btrfs_bio->csum = btrfs_bio->csum_inline;
2447                 icsum *= csum_size;
2448                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2449                        csum_size);
2450         }
2451
2452         bio_add_page(bio, page, failrec->len, pg_offset);
2453
2454         return bio;
2455 }
2456
2457 /*
2458  * this is a generic handler for readpage errors (default
2459  * readpage_io_failed_hook). if other copies exist, read those and write back
2460  * good data to the failed position. does not investigate in remapping the
2461  * failed extent elsewhere, hoping the device will be smart enough to do this as
2462  * needed
2463  */
2464
2465 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2466                               struct page *page, u64 start, u64 end,
2467                               int failed_mirror)
2468 {
2469         struct io_failure_record *failrec;
2470         struct inode *inode = page->mapping->host;
2471         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2472         struct bio *bio;
2473         int read_mode;
2474         int ret;
2475
2476         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2477
2478         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2479         if (ret)
2480                 return ret;
2481
2482         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2483         if (!ret) {
2484                 free_io_failure(inode, failrec);
2485                 return -EIO;
2486         }
2487
2488         if (failed_bio->bi_vcnt > 1)
2489                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2490         else
2491                 read_mode = READ_SYNC;
2492
2493         phy_offset >>= inode->i_sb->s_blocksize_bits;
2494         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2495                                       start - page_offset(page),
2496                                       (int)phy_offset, failed_bio->bi_end_io,
2497                                       NULL);
2498         if (!bio) {
2499                 free_io_failure(inode, failrec);
2500                 return -EIO;
2501         }
2502
2503         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2504                  read_mode, failrec->this_mirror, failrec->in_validation);
2505
2506         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2507                                          failrec->this_mirror,
2508                                          failrec->bio_flags, 0);
2509         if (ret) {
2510                 free_io_failure(inode, failrec);
2511                 bio_put(bio);
2512         }
2513
2514         return ret;
2515 }
2516
2517 /* lots and lots of room for performance fixes in the end_bio funcs */
2518
2519 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2520 {
2521         int uptodate = (err == 0);
2522         struct extent_io_tree *tree;
2523         int ret = 0;
2524
2525         tree = &BTRFS_I(page->mapping->host)->io_tree;
2526
2527         if (tree->ops && tree->ops->writepage_end_io_hook) {
2528                 ret = tree->ops->writepage_end_io_hook(page, start,
2529                                                end, NULL, uptodate);
2530                 if (ret)
2531                         uptodate = 0;
2532         }
2533
2534         if (!uptodate) {
2535                 ClearPageUptodate(page);
2536                 SetPageError(page);
2537                 ret = err < 0 ? err : -EIO;
2538                 mapping_set_error(page->mapping, ret);
2539         }
2540         return 0;
2541 }
2542
2543 /*
2544  * after a writepage IO is done, we need to:
2545  * clear the uptodate bits on error
2546  * clear the writeback bits in the extent tree for this IO
2547  * end_page_writeback if the page has no more pending IO
2548  *
2549  * Scheduling is not allowed, so the extent state tree is expected
2550  * to have one and only one object corresponding to this IO.
2551  */
2552 static void end_bio_extent_writepage(struct bio *bio)
2553 {
2554         struct bio_vec *bvec;
2555         u64 start;
2556         u64 end;
2557         int i;
2558
2559         bio_for_each_segment_all(bvec, bio, i) {
2560                 struct page *page = bvec->bv_page;
2561
2562                 /* We always issue full-page reads, but if some block
2563                  * in a page fails to read, blk_update_request() will
2564                  * advance bv_offset and adjust bv_len to compensate.
2565                  * Print a warning for nonzero offsets, and an error
2566                  * if they don't add up to a full page.  */
2567                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2568                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2569                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2570                                    "partial page write in btrfs with offset %u and length %u",
2571                                         bvec->bv_offset, bvec->bv_len);
2572                         else
2573                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2574                                    "incomplete page write in btrfs with offset %u and "
2575                                    "length %u",
2576                                         bvec->bv_offset, bvec->bv_len);
2577                 }
2578
2579                 start = page_offset(page);
2580                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2581
2582                 if (end_extent_writepage(page, bio->bi_error, start, end))
2583                         continue;
2584
2585                 end_page_writeback(page);
2586         }
2587
2588         bio_put(bio);
2589 }
2590
2591 static void
2592 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2593                               int uptodate)
2594 {
2595         struct extent_state *cached = NULL;
2596         u64 end = start + len - 1;
2597
2598         if (uptodate && tree->track_uptodate)
2599                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2600         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2601 }
2602
2603 /*
2604  * after a readpage IO is done, we need to:
2605  * clear the uptodate bits on error
2606  * set the uptodate bits if things worked
2607  * set the page up to date if all extents in the tree are uptodate
2608  * clear the lock bit in the extent tree
2609  * unlock the page if there are no other extents locked for it
2610  *
2611  * Scheduling is not allowed, so the extent state tree is expected
2612  * to have one and only one object corresponding to this IO.
2613  */
2614 static void end_bio_extent_readpage(struct bio *bio)
2615 {
2616         struct bio_vec *bvec;
2617         int uptodate = !bio->bi_error;
2618         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2619         struct extent_io_tree *tree;
2620         u64 offset = 0;
2621         u64 start;
2622         u64 end;
2623         u64 len;
2624         u64 extent_start = 0;
2625         u64 extent_len = 0;
2626         int mirror;
2627         int ret;
2628         int i;
2629
2630         bio_for_each_segment_all(bvec, bio, i) {
2631                 struct page *page = bvec->bv_page;
2632                 struct inode *inode = page->mapping->host;
2633
2634                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2635                          "mirror=%u\n", (u64)bio->bi_iter.bi_sector,
2636                          bio->bi_error, io_bio->mirror_num);
2637                 tree = &BTRFS_I(inode)->io_tree;
2638
2639                 /* We always issue full-page reads, but if some block
2640                  * in a page fails to read, blk_update_request() will
2641                  * advance bv_offset and adjust bv_len to compensate.
2642                  * Print a warning for nonzero offsets, and an error
2643                  * if they don't add up to a full page.  */
2644                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2645                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2646                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2647                                    "partial page read in btrfs with offset %u and length %u",
2648                                         bvec->bv_offset, bvec->bv_len);
2649                         else
2650                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2651                                    "incomplete page read in btrfs with offset %u and "
2652                                    "length %u",
2653                                         bvec->bv_offset, bvec->bv_len);
2654                 }
2655
2656                 start = page_offset(page);
2657                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2658                 len = bvec->bv_len;
2659
2660                 mirror = io_bio->mirror_num;
2661                 if (likely(uptodate && tree->ops &&
2662                            tree->ops->readpage_end_io_hook)) {
2663                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2664                                                               page, start, end,
2665                                                               mirror);
2666                         if (ret)
2667                                 uptodate = 0;
2668                         else
2669                                 clean_io_failure(inode, start, page, 0);
2670                 }
2671
2672                 if (likely(uptodate))
2673                         goto readpage_ok;
2674
2675                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2676                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2677                         if (!ret && !bio->bi_error)
2678                                 uptodate = 1;
2679                 } else {
2680                         /*
2681                          * The generic bio_readpage_error handles errors the
2682                          * following way: If possible, new read requests are
2683                          * created and submitted and will end up in
2684                          * end_bio_extent_readpage as well (if we're lucky, not
2685                          * in the !uptodate case). In that case it returns 0 and
2686                          * we just go on with the next page in our bio. If it
2687                          * can't handle the error it will return -EIO and we
2688                          * remain responsible for that page.
2689                          */
2690                         ret = bio_readpage_error(bio, offset, page, start, end,
2691                                                  mirror);
2692                         if (ret == 0) {
2693                                 uptodate = !bio->bi_error;
2694                                 offset += len;
2695                                 continue;
2696                         }
2697                 }
2698 readpage_ok:
2699                 if (likely(uptodate)) {
2700                         loff_t i_size = i_size_read(inode);
2701                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2702                         unsigned off;
2703
2704                         /* Zero out the end if this page straddles i_size */
2705                         off = i_size & (PAGE_CACHE_SIZE-1);
2706                         if (page->index == end_index && off)
2707                                 zero_user_segment(page, off, PAGE_CACHE_SIZE);
2708                         SetPageUptodate(page);
2709                 } else {
2710                         ClearPageUptodate(page);
2711                         SetPageError(page);
2712                 }
2713                 unlock_page(page);
2714                 offset += len;
2715
2716                 if (unlikely(!uptodate)) {
2717                         if (extent_len) {
2718                                 endio_readpage_release_extent(tree,
2719                                                               extent_start,
2720                                                               extent_len, 1);
2721                                 extent_start = 0;
2722                                 extent_len = 0;
2723                         }
2724                         endio_readpage_release_extent(tree, start,
2725                                                       end - start + 1, 0);
2726                 } else if (!extent_len) {
2727                         extent_start = start;
2728                         extent_len = end + 1 - start;
2729                 } else if (extent_start + extent_len == start) {
2730                         extent_len += end + 1 - start;
2731                 } else {
2732                         endio_readpage_release_extent(tree, extent_start,
2733                                                       extent_len, uptodate);
2734                         extent_start = start;
2735                         extent_len = end + 1 - start;
2736                 }
2737         }
2738
2739         if (extent_len)
2740                 endio_readpage_release_extent(tree, extent_start, extent_len,
2741                                               uptodate);
2742         if (io_bio->end_io)
2743                 io_bio->end_io(io_bio, bio->bi_error);
2744         bio_put(bio);
2745 }
2746
2747 /*
2748  * this allocates from the btrfs_bioset.  We're returning a bio right now
2749  * but you can call btrfs_io_bio for the appropriate container_of magic
2750  */
2751 struct bio *
2752 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2753                 gfp_t gfp_flags)
2754 {
2755         struct btrfs_io_bio *btrfs_bio;
2756         struct bio *bio;
2757
2758         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2759
2760         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2761                 while (!bio && (nr_vecs /= 2)) {
2762                         bio = bio_alloc_bioset(gfp_flags,
2763                                                nr_vecs, btrfs_bioset);
2764                 }
2765         }
2766
2767         if (bio) {
2768                 bio->bi_bdev = bdev;
2769                 bio->bi_iter.bi_sector = first_sector;
2770                 btrfs_bio = btrfs_io_bio(bio);
2771                 btrfs_bio->csum = NULL;
2772                 btrfs_bio->csum_allocated = NULL;
2773                 btrfs_bio->end_io = NULL;
2774         }
2775         return bio;
2776 }
2777
2778 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2779 {
2780         struct btrfs_io_bio *btrfs_bio;
2781         struct bio *new;
2782
2783         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2784         if (new) {
2785                 btrfs_bio = btrfs_io_bio(new);
2786                 btrfs_bio->csum = NULL;
2787                 btrfs_bio->csum_allocated = NULL;
2788                 btrfs_bio->end_io = NULL;
2789         }
2790         return new;
2791 }
2792
2793 /* this also allocates from the btrfs_bioset */
2794 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2795 {
2796         struct btrfs_io_bio *btrfs_bio;
2797         struct bio *bio;
2798
2799         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2800         if (bio) {
2801                 btrfs_bio = btrfs_io_bio(bio);
2802                 btrfs_bio->csum = NULL;
2803                 btrfs_bio->csum_allocated = NULL;
2804                 btrfs_bio->end_io = NULL;
2805         }
2806         return bio;
2807 }
2808
2809
2810 static int __must_check submit_one_bio(int rw, struct bio *bio,
2811                                        int mirror_num, unsigned long bio_flags)
2812 {
2813         int ret = 0;
2814         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2815         struct page *page = bvec->bv_page;
2816         struct extent_io_tree *tree = bio->bi_private;
2817         u64 start;
2818
2819         start = page_offset(page) + bvec->bv_offset;
2820
2821         bio->bi_private = NULL;
2822
2823         bio_get(bio);
2824
2825         if (tree->ops && tree->ops->submit_bio_hook)
2826                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2827                                            mirror_num, bio_flags, start);
2828         else
2829                 btrfsic_submit_bio(rw, bio);
2830
2831         bio_put(bio);
2832         return ret;
2833 }
2834
2835 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2836                      unsigned long offset, size_t size, struct bio *bio,
2837                      unsigned long bio_flags)
2838 {
2839         int ret = 0;
2840         if (tree->ops && tree->ops->merge_bio_hook)
2841                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2842                                                 bio_flags);
2843         BUG_ON(ret < 0);
2844         return ret;
2845
2846 }
2847
2848 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2849                               struct writeback_control *wbc,
2850                               struct page *page, sector_t sector,
2851                               size_t size, unsigned long offset,
2852                               struct block_device *bdev,
2853                               struct bio **bio_ret,
2854                               unsigned long max_pages,
2855                               bio_end_io_t end_io_func,
2856                               int mirror_num,
2857                               unsigned long prev_bio_flags,
2858                               unsigned long bio_flags,
2859                               bool force_bio_submit)
2860 {
2861         int ret = 0;
2862         struct bio *bio;
2863         int contig = 0;
2864         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2865         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2866
2867         if (bio_ret && *bio_ret) {
2868                 bio = *bio_ret;
2869                 if (old_compressed)
2870                         contig = bio->bi_iter.bi_sector == sector;
2871                 else
2872                         contig = bio_end_sector(bio) == sector;
2873
2874                 if (prev_bio_flags != bio_flags || !contig ||
2875                     force_bio_submit ||
2876                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2877                     bio_add_page(bio, page, page_size, offset) < page_size) {
2878                         ret = submit_one_bio(rw, bio, mirror_num,
2879                                              prev_bio_flags);
2880                         if (ret < 0) {
2881                                 *bio_ret = NULL;
2882                                 return ret;
2883                         }
2884                         bio = NULL;
2885                 } else {
2886                         if (wbc)
2887                                 wbc_account_io(wbc, page, page_size);
2888                         return 0;
2889                 }
2890         }
2891
2892         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2893                         GFP_NOFS | __GFP_HIGH);
2894         if (!bio)
2895                 return -ENOMEM;
2896
2897         bio_add_page(bio, page, page_size, offset);
2898         bio->bi_end_io = end_io_func;
2899         bio->bi_private = tree;
2900         if (wbc) {
2901                 wbc_init_bio(wbc, bio);
2902                 wbc_account_io(wbc, page, page_size);
2903         }
2904
2905         if (bio_ret)
2906                 *bio_ret = bio;
2907         else
2908                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2909
2910         return ret;
2911 }
2912
2913 static void attach_extent_buffer_page(struct extent_buffer *eb,
2914                                       struct page *page)
2915 {
2916         if (!PagePrivate(page)) {
2917                 SetPagePrivate(page);
2918                 page_cache_get(page);
2919                 set_page_private(page, (unsigned long)eb);
2920         } else {
2921                 WARN_ON(page->private != (unsigned long)eb);
2922         }
2923 }
2924
2925 void set_page_extent_mapped(struct page *page)
2926 {
2927         if (!PagePrivate(page)) {
2928                 SetPagePrivate(page);
2929                 page_cache_get(page);
2930                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2931         }
2932 }
2933
2934 static struct extent_map *
2935 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2936                  u64 start, u64 len, get_extent_t *get_extent,
2937                  struct extent_map **em_cached)
2938 {
2939         struct extent_map *em;
2940
2941         if (em_cached && *em_cached) {
2942                 em = *em_cached;
2943                 if (extent_map_in_tree(em) && start >= em->start &&
2944                     start < extent_map_end(em)) {
2945                         atomic_inc(&em->refs);
2946                         return em;
2947                 }
2948
2949                 free_extent_map(em);
2950                 *em_cached = NULL;
2951         }
2952
2953         em = get_extent(inode, page, pg_offset, start, len, 0);
2954         if (em_cached && !IS_ERR_OR_NULL(em)) {
2955                 BUG_ON(*em_cached);
2956                 atomic_inc(&em->refs);
2957                 *em_cached = em;
2958         }
2959         return em;
2960 }
2961 /*
2962  * basic readpage implementation.  Locked extent state structs are inserted
2963  * into the tree that are removed when the IO is done (by the end_io
2964  * handlers)
2965  * XXX JDM: This needs looking at to ensure proper page locking
2966  */
2967 static int __do_readpage(struct extent_io_tree *tree,
2968                          struct page *page,
2969                          get_extent_t *get_extent,
2970                          struct extent_map **em_cached,
2971                          struct bio **bio, int mirror_num,
2972                          unsigned long *bio_flags, int rw,
2973                          u64 *prev_em_start)
2974 {
2975         struct inode *inode = page->mapping->host;
2976         u64 start = page_offset(page);
2977         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2978         u64 end;
2979         u64 cur = start;
2980         u64 extent_offset;
2981         u64 last_byte = i_size_read(inode);
2982         u64 block_start;
2983         u64 cur_end;
2984         sector_t sector;
2985         struct extent_map *em;
2986         struct block_device *bdev;
2987         int ret;
2988         int nr = 0;
2989         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2990         size_t pg_offset = 0;
2991         size_t iosize;
2992         size_t disk_io_size;
2993         size_t blocksize = inode->i_sb->s_blocksize;
2994         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2995
2996         set_page_extent_mapped(page);
2997
2998         end = page_end;
2999         if (!PageUptodate(page)) {
3000                 if (cleancache_get_page(page) == 0) {
3001                         BUG_ON(blocksize != PAGE_SIZE);
3002                         unlock_extent(tree, start, end);
3003                         goto out;
3004                 }
3005         }
3006
3007         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
3008                 char *userpage;
3009                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
3010
3011                 if (zero_offset) {
3012                         iosize = PAGE_CACHE_SIZE - zero_offset;
3013                         userpage = kmap_atomic(page);
3014                         memset(userpage + zero_offset, 0, iosize);
3015                         flush_dcache_page(page);
3016                         kunmap_atomic(userpage);
3017                 }
3018         }
3019         while (cur <= end) {
3020                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
3021                 bool force_bio_submit = false;
3022
3023                 if (cur >= last_byte) {
3024                         char *userpage;
3025                         struct extent_state *cached = NULL;
3026
3027                         iosize = PAGE_CACHE_SIZE - pg_offset;
3028                         userpage = kmap_atomic(page);
3029                         memset(userpage + pg_offset, 0, iosize);
3030                         flush_dcache_page(page);
3031                         kunmap_atomic(userpage);
3032                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3033                                             &cached, GFP_NOFS);
3034                         if (!parent_locked)
3035                                 unlock_extent_cached(tree, cur,
3036                                                      cur + iosize - 1,
3037                                                      &cached, GFP_NOFS);
3038                         break;
3039                 }
3040                 em = __get_extent_map(inode, page, pg_offset, cur,
3041                                       end - cur + 1, get_extent, em_cached);
3042                 if (IS_ERR_OR_NULL(em)) {
3043                         SetPageError(page);
3044                         if (!parent_locked)
3045                                 unlock_extent(tree, cur, end);
3046                         break;
3047                 }
3048                 extent_offset = cur - em->start;
3049                 BUG_ON(extent_map_end(em) <= cur);
3050                 BUG_ON(end < cur);
3051
3052                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
3053                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
3054                         extent_set_compress_type(&this_bio_flag,
3055                                                  em->compress_type);
3056                 }
3057
3058                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
3059                 cur_end = min(extent_map_end(em) - 1, end);
3060                 iosize = ALIGN(iosize, blocksize);
3061                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
3062                         disk_io_size = em->block_len;
3063                         sector = em->block_start >> 9;
3064                 } else {
3065                         sector = (em->block_start + extent_offset) >> 9;
3066                         disk_io_size = iosize;
3067                 }
3068                 bdev = em->bdev;
3069                 block_start = em->block_start;
3070                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3071                         block_start = EXTENT_MAP_HOLE;
3072
3073                 /*
3074                  * If we have a file range that points to a compressed extent
3075                  * and it's followed by a consecutive file range that points to
3076                  * to the same compressed extent (possibly with a different
3077                  * offset and/or length, so it either points to the whole extent
3078                  * or only part of it), we must make sure we do not submit a
3079                  * single bio to populate the pages for the 2 ranges because
3080                  * this makes the compressed extent read zero out the pages
3081                  * belonging to the 2nd range. Imagine the following scenario:
3082                  *
3083                  *  File layout
3084                  *  [0 - 8K]                     [8K - 24K]
3085                  *    |                               |
3086                  *    |                               |
3087                  * points to extent X,         points to extent X,
3088                  * offset 4K, length of 8K     offset 0, length 16K
3089                  *
3090                  * [extent X, compressed length = 4K uncompressed length = 16K]
3091                  *
3092                  * If the bio to read the compressed extent covers both ranges,
3093                  * it will decompress extent X into the pages belonging to the
3094                  * first range and then it will stop, zeroing out the remaining
3095                  * pages that belong to the other range that points to extent X.
3096                  * So here we make sure we submit 2 bios, one for the first
3097                  * range and another one for the third range. Both will target
3098                  * the same physical extent from disk, but we can't currently
3099                  * make the compressed bio endio callback populate the pages
3100                  * for both ranges because each compressed bio is tightly
3101                  * coupled with a single extent map, and each range can have
3102                  * an extent map with a different offset value relative to the
3103                  * uncompressed data of our extent and different lengths. This
3104                  * is a corner case so we prioritize correctness over
3105                  * non-optimal behavior (submitting 2 bios for the same extent).
3106                  */
3107                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3108                     prev_em_start && *prev_em_start != (u64)-1 &&
3109                     *prev_em_start != em->start)
3110                         force_bio_submit = true;
3111
3112                 if (prev_em_start)
3113                         *prev_em_start = em->start;
3114
3115                 free_extent_map(em);
3116                 em = NULL;
3117
3118                 /* we've found a hole, just zero and go on */
3119                 if (block_start == EXTENT_MAP_HOLE) {
3120                         char *userpage;
3121                         struct extent_state *cached = NULL;
3122
3123                         userpage = kmap_atomic(page);
3124                         memset(userpage + pg_offset, 0, iosize);
3125                         flush_dcache_page(page);
3126                         kunmap_atomic(userpage);
3127
3128                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3129                                             &cached, GFP_NOFS);
3130                         if (parent_locked)
3131                                 free_extent_state(cached);
3132                         else
3133                                 unlock_extent_cached(tree, cur,
3134                                                      cur + iosize - 1,
3135                                                      &cached, GFP_NOFS);
3136                         cur = cur + iosize;
3137                         pg_offset += iosize;
3138                         continue;
3139                 }
3140                 /* the get_extent function already copied into the page */
3141                 if (test_range_bit(tree, cur, cur_end,
3142                                    EXTENT_UPTODATE, 1, NULL)) {
3143                         check_page_uptodate(tree, page);
3144                         if (!parent_locked)
3145                                 unlock_extent(tree, cur, cur + iosize - 1);
3146                         cur = cur + iosize;
3147                         pg_offset += iosize;
3148                         continue;
3149                 }
3150                 /* we have an inline extent but it didn't get marked up
3151                  * to date.  Error out
3152                  */
3153                 if (block_start == EXTENT_MAP_INLINE) {
3154                         SetPageError(page);
3155                         if (!parent_locked)
3156                                 unlock_extent(tree, cur, cur + iosize - 1);
3157                         cur = cur + iosize;
3158                         pg_offset += iosize;
3159                         continue;
3160                 }
3161
3162                 pnr -= page->index;
3163                 ret = submit_extent_page(rw, tree, NULL, page,
3164                                          sector, disk_io_size, pg_offset,
3165                                          bdev, bio, pnr,
3166                                          end_bio_extent_readpage, mirror_num,
3167                                          *bio_flags,
3168                                          this_bio_flag,
3169                                          force_bio_submit);
3170                 if (!ret) {
3171                         nr++;
3172                         *bio_flags = this_bio_flag;
3173                 } else {
3174                         SetPageError(page);
3175                         if (!parent_locked)
3176                                 unlock_extent(tree, cur, cur + iosize - 1);
3177                 }
3178                 cur = cur + iosize;
3179                 pg_offset += iosize;
3180         }
3181 out:
3182         if (!nr) {
3183                 if (!PageError(page))
3184                         SetPageUptodate(page);
3185                 unlock_page(page);
3186         }
3187         return 0;
3188 }
3189
3190 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3191                                              struct page *pages[], int nr_pages,
3192                                              u64 start, u64 end,
3193                                              get_extent_t *get_extent,
3194                                              struct extent_map **em_cached,
3195                                              struct bio **bio, int mirror_num,
3196                                              unsigned long *bio_flags, int rw,
3197                                              u64 *prev_em_start)
3198 {
3199         struct inode *inode;
3200         struct btrfs_ordered_extent *ordered;
3201         int index;
3202
3203         inode = pages[0]->mapping->host;
3204         while (1) {
3205                 lock_extent(tree, start, end);
3206                 ordered = btrfs_lookup_ordered_range(inode, start,
3207                                                      end - start + 1);
3208                 if (!ordered)
3209                         break;
3210                 unlock_extent(tree, start, end);
3211                 btrfs_start_ordered_extent(inode, ordered, 1);
3212                 btrfs_put_ordered_extent(ordered);
3213         }
3214
3215         for (index = 0; index < nr_pages; index++) {
3216                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3217                               mirror_num, bio_flags, rw, prev_em_start);
3218                 page_cache_release(pages[index]);
3219         }
3220 }
3221
3222 static void __extent_readpages(struct extent_io_tree *tree,
3223                                struct page *pages[],
3224                                int nr_pages, get_extent_t *get_extent,
3225                                struct extent_map **em_cached,
3226                                struct bio **bio, int mirror_num,
3227                                unsigned long *bio_flags, int rw,
3228                                u64 *prev_em_start)
3229 {
3230         u64 start = 0;
3231         u64 end = 0;
3232         u64 page_start;
3233         int index;
3234         int first_index = 0;
3235
3236         for (index = 0; index < nr_pages; index++) {
3237                 page_start = page_offset(pages[index]);
3238                 if (!end) {
3239                         start = page_start;
3240                         end = start + PAGE_CACHE_SIZE - 1;
3241                         first_index = index;
3242                 } else if (end + 1 == page_start) {
3243                         end += PAGE_CACHE_SIZE;
3244                 } else {
3245                         __do_contiguous_readpages(tree, &pages[first_index],
3246                                                   index - first_index, start,
3247                                                   end, get_extent, em_cached,
3248                                                   bio, mirror_num, bio_flags,
3249                                                   rw, prev_em_start);
3250                         start = page_start;
3251                         end = start + PAGE_CACHE_SIZE - 1;
3252                         first_index = index;
3253                 }
3254         }
3255
3256         if (end)
3257                 __do_contiguous_readpages(tree, &pages[first_index],
3258                                           index - first_index, start,
3259                                           end, get_extent, em_cached, bio,
3260                                           mirror_num, bio_flags, rw,
3261                                           prev_em_start);
3262 }
3263
3264 static int __extent_read_full_page(struct extent_io_tree *tree,
3265                                    struct page *page,
3266                                    get_extent_t *get_extent,
3267                                    struct bio **bio, int mirror_num,
3268                                    unsigned long *bio_flags, int rw)
3269 {
3270         struct inode *inode = page->mapping->host;
3271         struct btrfs_ordered_extent *ordered;
3272         u64 start = page_offset(page);
3273         u64 end = start + PAGE_CACHE_SIZE - 1;
3274         int ret;
3275
3276         while (1) {
3277                 lock_extent(tree, start, end);
3278                 ordered = btrfs_lookup_ordered_extent(inode, start);
3279                 if (!ordered)
3280                         break;
3281                 unlock_extent(tree, start, end);
3282                 btrfs_start_ordered_extent(inode, ordered, 1);
3283                 btrfs_put_ordered_extent(ordered);
3284         }
3285
3286         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3287                             bio_flags, rw, NULL);
3288         return ret;
3289 }
3290
3291 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3292                             get_extent_t *get_extent, int mirror_num)
3293 {
3294         struct bio *bio = NULL;
3295         unsigned long bio_flags = 0;
3296         int ret;
3297
3298         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3299                                       &bio_flags, READ);
3300         if (bio)
3301                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3302         return ret;
3303 }
3304
3305 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3306                                  get_extent_t *get_extent, int mirror_num)
3307 {
3308         struct bio *bio = NULL;
3309         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3310         int ret;
3311
3312         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3313                             &bio_flags, READ, NULL);
3314         if (bio)
3315                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3316         return ret;
3317 }
3318
3319 static noinline void update_nr_written(struct page *page,
3320                                       struct writeback_control *wbc,
3321                                       unsigned long nr_written)
3322 {
3323         wbc->nr_to_write -= nr_written;
3324         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3325             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3326                 page->mapping->writeback_index = page->index + nr_written;
3327 }
3328
3329 /*
3330  * helper for __extent_writepage, doing all of the delayed allocation setup.
3331  *
3332  * This returns 1 if our fill_delalloc function did all the work required
3333  * to write the page (copy into inline extent).  In this case the IO has
3334  * been started and the page is already unlocked.
3335  *
3336  * This returns 0 if all went well (page still locked)
3337  * This returns < 0 if there were errors (page still locked)
3338  */
3339 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3340                               struct page *page, struct writeback_control *wbc,
3341                               struct extent_page_data *epd,
3342                               u64 delalloc_start,
3343                               unsigned long *nr_written)
3344 {
3345         struct extent_io_tree *tree = epd->tree;
3346         u64 page_end = delalloc_start + PAGE_CACHE_SIZE - 1;
3347         u64 nr_delalloc;
3348         u64 delalloc_to_write = 0;
3349         u64 delalloc_end = 0;
3350         int ret;
3351         int page_started = 0;
3352
3353         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3354                 return 0;
3355
3356         while (delalloc_end < page_end) {
3357                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3358                                                page,
3359                                                &delalloc_start,
3360                                                &delalloc_end,
3361                                                BTRFS_MAX_EXTENT_SIZE);
3362                 if (nr_delalloc == 0) {
3363                         delalloc_start = delalloc_end + 1;
3364                         continue;
3365                 }
3366                 ret = tree->ops->fill_delalloc(inode, page,
3367                                                delalloc_start,
3368                                                delalloc_end,
3369                                                &page_started,
3370                                                nr_written);
3371                 /* File system has been set read-only */
3372                 if (ret) {
3373                         SetPageError(page);
3374                         /* fill_delalloc should be return < 0 for error
3375                          * but just in case, we use > 0 here meaning the
3376                          * IO is started, so we don't want to return > 0
3377                          * unless things are going well.
3378                          */
3379                         ret = ret < 0 ? ret : -EIO;
3380                         goto done;
3381                 }
3382                 /*
3383                  * delalloc_end is already one less than the total
3384                  * length, so we don't subtract one from
3385                  * PAGE_CACHE_SIZE
3386                  */
3387                 delalloc_to_write += (delalloc_end - delalloc_start +
3388                                       PAGE_CACHE_SIZE) >>
3389                                       PAGE_CACHE_SHIFT;
3390                 delalloc_start = delalloc_end + 1;
3391         }
3392         if (wbc->nr_to_write < delalloc_to_write) {
3393                 int thresh = 8192;
3394
3395                 if (delalloc_to_write < thresh * 2)
3396                         thresh = delalloc_to_write;
3397                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3398                                          thresh);
3399         }
3400
3401         /* did the fill delalloc function already unlock and start
3402          * the IO?
3403          */
3404         if (page_started) {
3405                 /*
3406                  * we've unlocked the page, so we can't update
3407                  * the mapping's writeback index, just update
3408                  * nr_to_write.
3409                  */
3410                 wbc->nr_to_write -= *nr_written;
3411                 return 1;
3412         }
3413
3414         ret = 0;
3415
3416 done:
3417         return ret;
3418 }
3419
3420 /*
3421  * helper for __extent_writepage.  This calls the writepage start hooks,
3422  * and does the loop to map the page into extents and bios.
3423  *
3424  * We return 1 if the IO is started and the page is unlocked,
3425  * 0 if all went well (page still locked)
3426  * < 0 if there were errors (page still locked)
3427  */
3428 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3429                                  struct page *page,
3430                                  struct writeback_control *wbc,
3431                                  struct extent_page_data *epd,
3432                                  loff_t i_size,
3433                                  unsigned long nr_written,
3434                                  int write_flags, int *nr_ret)
3435 {
3436         struct extent_io_tree *tree = epd->tree;
3437         u64 start = page_offset(page);
3438         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3439         u64 end;
3440         u64 cur = start;
3441         u64 extent_offset;
3442         u64 block_start;
3443         u64 iosize;
3444         sector_t sector;
3445         struct extent_state *cached_state = NULL;
3446         struct extent_map *em;
3447         struct block_device *bdev;
3448         size_t pg_offset = 0;
3449         size_t blocksize;
3450         int ret = 0;
3451         int nr = 0;
3452         bool compressed;
3453
3454         if (tree->ops && tree->ops->writepage_start_hook) {
3455                 ret = tree->ops->writepage_start_hook(page, start,
3456                                                       page_end);
3457                 if (ret) {
3458                         /* Fixup worker will requeue */
3459                         if (ret == -EBUSY)
3460                                 wbc->pages_skipped++;
3461                         else
3462                                 redirty_page_for_writepage(wbc, page);
3463
3464                         update_nr_written(page, wbc, nr_written);
3465                         unlock_page(page);
3466                         ret = 1;
3467                         goto done_unlocked;
3468                 }
3469         }
3470
3471         /*
3472          * we don't want to touch the inode after unlocking the page,
3473          * so we update the mapping writeback index now
3474          */
3475         update_nr_written(page, wbc, nr_written + 1);
3476
3477         end = page_end;
3478         if (i_size <= start) {
3479                 if (tree->ops && tree->ops->writepage_end_io_hook)
3480                         tree->ops->writepage_end_io_hook(page, start,
3481                                                          page_end, NULL, 1);
3482                 goto done;
3483         }
3484
3485         blocksize = inode->i_sb->s_blocksize;
3486
3487         while (cur <= end) {
3488                 u64 em_end;
3489                 if (cur >= i_size) {
3490                         if (tree->ops && tree->ops->writepage_end_io_hook)
3491                                 tree->ops->writepage_end_io_hook(page, cur,
3492                                                          page_end, NULL, 1);
3493                         break;
3494                 }
3495                 em = epd->get_extent(inode, page, pg_offset, cur,
3496                                      end - cur + 1, 1);
3497                 if (IS_ERR_OR_NULL(em)) {
3498                         SetPageError(page);
3499                         ret = PTR_ERR_OR_ZERO(em);
3500                         break;
3501                 }
3502
3503                 extent_offset = cur - em->start;
3504                 em_end = extent_map_end(em);
3505                 BUG_ON(em_end <= cur);
3506                 BUG_ON(end < cur);
3507                 iosize = min(em_end - cur, end - cur + 1);
3508                 iosize = ALIGN(iosize, blocksize);
3509                 sector = (em->block_start + extent_offset) >> 9;
3510                 bdev = em->bdev;
3511                 block_start = em->block_start;
3512                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3513                 free_extent_map(em);
3514                 em = NULL;
3515
3516                 /*
3517                  * compressed and inline extents are written through other
3518                  * paths in the FS
3519                  */
3520                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3521                     block_start == EXTENT_MAP_INLINE) {
3522                         /*
3523                          * end_io notification does not happen here for
3524                          * compressed extents
3525                          */
3526                         if (!compressed && tree->ops &&
3527                             tree->ops->writepage_end_io_hook)
3528                                 tree->ops->writepage_end_io_hook(page, cur,
3529                                                          cur + iosize - 1,
3530                                                          NULL, 1);
3531                         else if (compressed) {
3532                                 /* we don't want to end_page_writeback on
3533                                  * a compressed extent.  this happens
3534                                  * elsewhere
3535                                  */
3536                                 nr++;
3537                         }
3538
3539                         cur += iosize;
3540                         pg_offset += iosize;
3541                         continue;
3542                 }
3543
3544                 if (tree->ops && tree->ops->writepage_io_hook) {
3545                         ret = tree->ops->writepage_io_hook(page, cur,
3546                                                 cur + iosize - 1);
3547                 } else {
3548                         ret = 0;
3549                 }
3550                 if (ret) {
3551                         SetPageError(page);
3552                 } else {
3553                         unsigned long max_nr = (i_size >> PAGE_CACHE_SHIFT) + 1;
3554
3555                         set_range_writeback(tree, cur, cur + iosize - 1);
3556                         if (!PageWriteback(page)) {
3557                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
3558                                            "page %lu not writeback, cur %llu end %llu",
3559                                        page->index, cur, end);
3560                         }
3561
3562                         ret = submit_extent_page(write_flags, tree, wbc, page,
3563                                                  sector, iosize, pg_offset,
3564                                                  bdev, &epd->bio, max_nr,
3565                                                  end_bio_extent_writepage,
3566                                                  0, 0, 0, false);
3567                         if (ret)
3568                                 SetPageError(page);
3569                 }
3570                 cur = cur + iosize;
3571                 pg_offset += iosize;
3572                 nr++;
3573         }
3574 done:
3575         *nr_ret = nr;
3576
3577 done_unlocked:
3578
3579         /* drop our reference on any cached states */
3580         free_extent_state(cached_state);
3581         return ret;
3582 }
3583
3584 /*
3585  * the writepage semantics are similar to regular writepage.  extent
3586  * records are inserted to lock ranges in the tree, and as dirty areas
3587  * are found, they are marked writeback.  Then the lock bits are removed
3588  * and the end_io handler clears the writeback ranges
3589  */
3590 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3591                               void *data)
3592 {
3593         struct inode *inode = page->mapping->host;
3594         struct extent_page_data *epd = data;
3595         u64 start = page_offset(page);
3596         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3597         int ret;
3598         int nr = 0;
3599         size_t pg_offset = 0;
3600         loff_t i_size = i_size_read(inode);
3601         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3602         int write_flags;
3603         unsigned long nr_written = 0;
3604
3605         if (wbc->sync_mode == WB_SYNC_ALL)
3606                 write_flags = WRITE_SYNC;
3607         else
3608                 write_flags = WRITE;
3609
3610         trace___extent_writepage(page, inode, wbc);
3611
3612         WARN_ON(!PageLocked(page));
3613
3614         ClearPageError(page);
3615
3616         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3617         if (page->index > end_index ||
3618            (page->index == end_index && !pg_offset)) {
3619                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3620                 unlock_page(page);
3621                 return 0;
3622         }
3623
3624         if (page->index == end_index) {
3625                 char *userpage;
3626
3627                 userpage = kmap_atomic(page);
3628                 memset(userpage + pg_offset, 0,
3629                        PAGE_CACHE_SIZE - pg_offset);
3630                 kunmap_atomic(userpage);
3631                 flush_dcache_page(page);
3632         }
3633
3634         pg_offset = 0;
3635
3636         set_page_extent_mapped(page);
3637
3638         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3639         if (ret == 1)
3640                 goto done_unlocked;
3641         if (ret)
3642                 goto done;
3643
3644         ret = __extent_writepage_io(inode, page, wbc, epd,
3645                                     i_size, nr_written, write_flags, &nr);
3646         if (ret == 1)
3647                 goto done_unlocked;
3648
3649 done:
3650         if (nr == 0) {
3651                 /* make sure the mapping tag for page dirty gets cleared */
3652                 set_page_writeback(page);
3653                 end_page_writeback(page);
3654         }
3655         if (PageError(page)) {
3656                 ret = ret < 0 ? ret : -EIO;
3657                 end_extent_writepage(page, ret, start, page_end);
3658         }
3659         unlock_page(page);
3660         return ret;
3661
3662 done_unlocked:
3663         return 0;
3664 }
3665
3666 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3667 {
3668         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3669                        TASK_UNINTERRUPTIBLE);
3670 }
3671
3672 static noinline_for_stack int
3673 lock_extent_buffer_for_io(struct extent_buffer *eb,
3674                           struct btrfs_fs_info *fs_info,
3675                           struct extent_page_data *epd)
3676 {
3677         unsigned long i, num_pages;
3678         int flush = 0;
3679         int ret = 0;
3680
3681         if (!btrfs_try_tree_write_lock(eb)) {
3682                 flush = 1;
3683                 flush_write_bio(epd);
3684                 btrfs_tree_lock(eb);
3685         }
3686
3687         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3688                 btrfs_tree_unlock(eb);
3689                 if (!epd->sync_io)
3690                         return 0;
3691                 if (!flush) {
3692                         flush_write_bio(epd);
3693                         flush = 1;
3694                 }
3695                 while (1) {
3696                         wait_on_extent_buffer_writeback(eb);
3697                         btrfs_tree_lock(eb);
3698                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3699                                 break;
3700                         btrfs_tree_unlock(eb);
3701                 }
3702         }
3703
3704         /*
3705          * We need to do this to prevent races in people who check if the eb is
3706          * under IO since we can end up having no IO bits set for a short period
3707          * of time.
3708          */
3709         spin_lock(&eb->refs_lock);
3710         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3711                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3712                 spin_unlock(&eb->refs_lock);
3713                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3714                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3715                                      -eb->len,
3716                                      fs_info->dirty_metadata_batch);
3717                 ret = 1;
3718         } else {
3719                 spin_unlock(&eb->refs_lock);
3720         }
3721
3722         btrfs_tree_unlock(eb);
3723
3724         if (!ret)
3725                 return ret;
3726
3727         num_pages = num_extent_pages(eb->start, eb->len);
3728         for (i = 0; i < num_pages; i++) {
3729                 struct page *p = eb->pages[i];
3730
3731                 if (!trylock_page(p)) {
3732                         if (!flush) {
3733                                 flush_write_bio(epd);
3734                                 flush = 1;
3735                         }
3736                         lock_page(p);
3737                 }
3738         }
3739
3740         return ret;
3741 }
3742
3743 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3744 {
3745         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3746         smp_mb__after_atomic();
3747         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3748 }
3749
3750 static void set_btree_ioerr(struct page *page)
3751 {
3752         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3753         struct btrfs_inode *btree_ino = BTRFS_I(eb->fs_info->btree_inode);
3754
3755         SetPageError(page);
3756         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3757                 return;
3758
3759         /*
3760          * If writeback for a btree extent that doesn't belong to a log tree
3761          * failed, increment the counter transaction->eb_write_errors.
3762          * We do this because while the transaction is running and before it's
3763          * committing (when we call filemap_fdata[write|wait]_range against
3764          * the btree inode), we might have
3765          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3766          * returns an error or an error happens during writeback, when we're
3767          * committing the transaction we wouldn't know about it, since the pages
3768          * can be no longer dirty nor marked anymore for writeback (if a
3769          * subsequent modification to the extent buffer didn't happen before the
3770          * transaction commit), which makes filemap_fdata[write|wait]_range not
3771          * able to find the pages tagged with SetPageError at transaction
3772          * commit time. So if this happens we must abort the transaction,
3773          * otherwise we commit a super block with btree roots that point to
3774          * btree nodes/leafs whose content on disk is invalid - either garbage
3775          * or the content of some node/leaf from a past generation that got
3776          * cowed or deleted and is no longer valid.
3777          *
3778          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3779          * not be enough - we need to distinguish between log tree extents vs
3780          * non-log tree extents, and the next filemap_fdatawait_range() call
3781          * will catch and clear such errors in the mapping - and that call might
3782          * be from a log sync and not from a transaction commit. Also, checking
3783          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3784          * not done and would not be reliable - the eb might have been released
3785          * from memory and reading it back again means that flag would not be
3786          * set (since it's a runtime flag, not persisted on disk).
3787          *
3788          * Using the flags below in the btree inode also makes us achieve the
3789          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3790          * writeback for all dirty pages and before filemap_fdatawait_range()
3791          * is called, the writeback for all dirty pages had already finished
3792          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3793          * filemap_fdatawait_range() would return success, as it could not know
3794          * that writeback errors happened (the pages were no longer tagged for
3795          * writeback).
3796          */
3797         switch (eb->log_index) {
3798         case -1:
3799                 set_bit(BTRFS_INODE_BTREE_ERR, &btree_ino->runtime_flags);
3800                 break;
3801         case 0:
3802                 set_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
3803                 break;
3804         case 1:
3805                 set_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
3806                 break;
3807         default:
3808                 BUG(); /* unexpected, logic error */
3809         }
3810 }
3811
3812 static void end_bio_extent_buffer_writepage(struct bio *bio)
3813 {
3814         struct bio_vec *bvec;
3815         struct extent_buffer *eb;
3816         int i, done;
3817
3818         bio_for_each_segment_all(bvec, bio, i) {
3819                 struct page *page = bvec->bv_page;
3820
3821                 eb = (struct extent_buffer *)page->private;
3822                 BUG_ON(!eb);
3823                 done = atomic_dec_and_test(&eb->io_pages);
3824
3825                 if (bio->bi_error ||
3826                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3827                         ClearPageUptodate(page);
3828                         set_btree_ioerr(page);
3829                 }
3830
3831                 end_page_writeback(page);
3832
3833                 if (!done)
3834                         continue;
3835
3836                 end_extent_buffer_writeback(eb);
3837         }
3838
3839         bio_put(bio);
3840 }
3841
3842 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3843                         struct btrfs_fs_info *fs_info,
3844                         struct writeback_control *wbc,
3845                         struct extent_page_data *epd)
3846 {
3847         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3848         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3849         u64 offset = eb->start;
3850         u32 nritems;
3851         unsigned long i, num_pages;
3852         unsigned long bio_flags = 0;
3853         unsigned long start, end;
3854         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3855         int ret = 0;
3856
3857         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3858         num_pages = num_extent_pages(eb->start, eb->len);
3859         atomic_set(&eb->io_pages, num_pages);
3860         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3861                 bio_flags = EXTENT_BIO_TREE_LOG;
3862
3863         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3864         nritems = btrfs_header_nritems(eb);
3865         if (btrfs_header_level(eb) > 0) {
3866                 end = btrfs_node_key_ptr_offset(nritems);
3867
3868                 memset_extent_buffer(eb, 0, end, eb->len - end);
3869         } else {
3870                 /*
3871                  * leaf:
3872                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3873                  */
3874                 start = btrfs_item_nr_offset(nritems);
3875                 end = btrfs_leaf_data(eb) +
3876                       leaf_data_end(fs_info->tree_root, eb);
3877                 memset_extent_buffer(eb, 0, start, end - start);
3878         }
3879
3880         for (i = 0; i < num_pages; i++) {
3881                 struct page *p = eb->pages[i];
3882
3883                 clear_page_dirty_for_io(p);
3884                 set_page_writeback(p);
3885                 ret = submit_extent_page(rw, tree, wbc, p, offset >> 9,
3886                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3887                                          -1, end_bio_extent_buffer_writepage,
3888                                          0, epd->bio_flags, bio_flags, false);
3889                 epd->bio_flags = bio_flags;
3890                 if (ret) {
3891                         set_btree_ioerr(p);
3892                         end_page_writeback(p);
3893                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3894                                 end_extent_buffer_writeback(eb);
3895                         ret = -EIO;
3896                         break;
3897                 }
3898                 offset += PAGE_CACHE_SIZE;
3899                 update_nr_written(p, wbc, 1);
3900                 unlock_page(p);
3901         }
3902
3903         if (unlikely(ret)) {
3904                 for (; i < num_pages; i++) {
3905                         struct page *p = eb->pages[i];
3906                         clear_page_dirty_for_io(p);
3907                         unlock_page(p);
3908                 }
3909         }
3910
3911         return ret;
3912 }
3913
3914 int btree_write_cache_pages(struct address_space *mapping,
3915                                    struct writeback_control *wbc)
3916 {
3917         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3918         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3919         struct extent_buffer *eb, *prev_eb = NULL;
3920         struct extent_page_data epd = {
3921                 .bio = NULL,
3922                 .tree = tree,
3923                 .extent_locked = 0,
3924                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3925                 .bio_flags = 0,
3926         };
3927         int ret = 0;
3928         int done = 0;
3929         int nr_to_write_done = 0;
3930         struct pagevec pvec;
3931         int nr_pages;
3932         pgoff_t index;
3933         pgoff_t end;            /* Inclusive */
3934         int scanned = 0;
3935         int tag;
3936
3937         pagevec_init(&pvec, 0);
3938         if (wbc->range_cyclic) {
3939                 index = mapping->writeback_index; /* Start from prev offset */
3940                 end = -1;
3941         } else {
3942                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3943                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3944                 scanned = 1;
3945         }
3946         if (wbc->sync_mode == WB_SYNC_ALL)
3947                 tag = PAGECACHE_TAG_TOWRITE;
3948         else
3949                 tag = PAGECACHE_TAG_DIRTY;
3950 retry:
3951         if (wbc->sync_mode == WB_SYNC_ALL)
3952                 tag_pages_for_writeback(mapping, index, end);
3953         while (!done && !nr_to_write_done && (index <= end) &&
3954                (nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
3955                         tag))) {
3956                 unsigned i;
3957
3958                 scanned = 1;
3959                 for (i = 0; i < nr_pages; i++) {
3960                         struct page *page = pvec.pages[i];
3961
3962                         if (!PagePrivate(page))
3963                                 continue;
3964
3965                         spin_lock(&mapping->private_lock);
3966                         if (!PagePrivate(page)) {
3967                                 spin_unlock(&mapping->private_lock);
3968                                 continue;
3969                         }
3970
3971                         eb = (struct extent_buffer *)page->private;
3972
3973                         /*
3974                          * Shouldn't happen and normally this would be a BUG_ON
3975                          * but no sense in crashing the users box for something
3976                          * we can survive anyway.
3977                          */
3978                         if (WARN_ON(!eb)) {
3979                                 spin_unlock(&mapping->private_lock);
3980                                 continue;
3981                         }
3982
3983                         if (eb == prev_eb) {
3984                                 spin_unlock(&mapping->private_lock);
3985                                 continue;
3986                         }
3987
3988                         ret = atomic_inc_not_zero(&eb->refs);
3989                         spin_unlock(&mapping->private_lock);
3990                         if (!ret)
3991                                 continue;
3992
3993                         prev_eb = eb;
3994                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3995                         if (!ret) {
3996                                 free_extent_buffer(eb);
3997                                 continue;
3998                         }
3999
4000                         ret = write_one_eb(eb, fs_info, wbc, &epd);
4001                         if (ret) {
4002                                 done = 1;
4003                                 free_extent_buffer(eb);
4004                                 break;
4005                         }
4006                         free_extent_buffer(eb);
4007
4008                         /*
4009                          * the filesystem may choose to bump up nr_to_write.
4010                          * We have to make sure to honor the new nr_to_write
4011                          * at any time
4012                          */
4013                         nr_to_write_done = wbc->nr_to_write <= 0;
4014                 }
4015                 pagevec_release(&pvec);
4016                 cond_resched();
4017         }
4018         if (!scanned && !done) {
4019                 /*
4020                  * We hit the last page and there is more work to be done: wrap
4021                  * back to the start of the file
4022                  */
4023                 scanned = 1;
4024                 index = 0;
4025                 goto retry;
4026         }
4027         flush_write_bio(&epd);
4028         return ret;
4029 }
4030
4031 /**
4032  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
4033  * @mapping: address space structure to write
4034  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
4035  * @writepage: function called for each page
4036  * @data: data passed to writepage function
4037  *
4038  * If a page is already under I/O, write_cache_pages() skips it, even
4039  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
4040  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
4041  * and msync() need to guarantee that all the data which was dirty at the time
4042  * the call was made get new I/O started against them.  If wbc->sync_mode is
4043  * WB_SYNC_ALL then we were called for data integrity and we must wait for
4044  * existing IO to complete.
4045  */
4046 static int extent_write_cache_pages(struct extent_io_tree *tree,
4047                              struct address_space *mapping,
4048                              struct writeback_control *wbc,
4049                              writepage_t writepage, void *data,
4050                              void (*flush_fn)(void *))
4051 {
4052         struct inode *inode = mapping->host;
4053         int ret = 0;
4054         int done = 0;
4055         int err = 0;
4056         int nr_to_write_done = 0;
4057         struct pagevec pvec;
4058         int nr_pages;
4059         pgoff_t index;
4060         pgoff_t end;            /* Inclusive */
4061         int scanned = 0;
4062         int tag;
4063
4064         /*
4065          * We have to hold onto the inode so that ordered extents can do their
4066          * work when the IO finishes.  The alternative to this is failing to add
4067          * an ordered extent if the igrab() fails there and that is a huge pain
4068          * to deal with, so instead just hold onto the inode throughout the
4069          * writepages operation.  If it fails here we are freeing up the inode
4070          * anyway and we'd rather not waste our time writing out stuff that is
4071          * going to be truncated anyway.
4072          */
4073         if (!igrab(inode))
4074                 return 0;
4075
4076         pagevec_init(&pvec, 0);
4077         if (wbc->range_cyclic) {
4078                 index = mapping->writeback_index; /* Start from prev offset */
4079                 end = -1;
4080         } else {
4081                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
4082                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
4083                 scanned = 1;
4084         }
4085         if (wbc->sync_mode == WB_SYNC_ALL)
4086                 tag = PAGECACHE_TAG_TOWRITE;
4087         else
4088                 tag = PAGECACHE_TAG_DIRTY;
4089 retry:
4090         if (wbc->sync_mode == WB_SYNC_ALL)
4091                 tag_pages_for_writeback(mapping, index, end);
4092         while (!done && !nr_to_write_done && (index <= end) &&
4093                         (nr_pages = pagevec_lookup_range_tag(&pvec, mapping,
4094                                                 &index, end, tag))) {
4095                 unsigned i;
4096
4097                 scanned = 1;
4098                 for (i = 0; i < nr_pages; i++) {
4099                         struct page *page = pvec.pages[i];
4100
4101                         /*
4102                          * At this point we hold neither mapping->tree_lock nor
4103                          * lock on the page itself: the page may be truncated or
4104                          * invalidated (changing page->mapping to NULL), or even
4105                          * swizzled back from swapper_space to tmpfs file
4106                          * mapping
4107                          */
4108                         if (!trylock_page(page)) {
4109                                 flush_fn(data);
4110                                 lock_page(page);
4111                         }
4112
4113                         if (unlikely(page->mapping != mapping)) {
4114                                 unlock_page(page);
4115                                 continue;
4116                         }
4117
4118                         if (wbc->sync_mode != WB_SYNC_NONE) {
4119                                 if (PageWriteback(page))
4120                                         flush_fn(data);
4121                                 wait_on_page_writeback(page);
4122                         }
4123
4124                         if (PageWriteback(page) ||
4125                             !clear_page_dirty_for_io(page)) {
4126                                 unlock_page(page);
4127                                 continue;
4128                         }
4129
4130                         ret = (*writepage)(page, wbc, data);
4131
4132                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4133                                 unlock_page(page);
4134                                 ret = 0;
4135                         }
4136                         if (!err && ret < 0)
4137                                 err = ret;
4138
4139                         /*
4140                          * the filesystem may choose to bump up nr_to_write.
4141                          * We have to make sure to honor the new nr_to_write
4142                          * at any time
4143                          */
4144                         nr_to_write_done = wbc->nr_to_write <= 0;
4145                 }
4146                 pagevec_release(&pvec);
4147                 cond_resched();
4148         }
4149         if (!scanned && !done && !err) {
4150                 /*
4151                  * We hit the last page and there is more work to be done: wrap
4152                  * back to the start of the file
4153                  */
4154                 scanned = 1;
4155                 index = 0;
4156                 goto retry;
4157         }
4158         btrfs_add_delayed_iput(inode);
4159         return err;
4160 }
4161
4162 static void flush_epd_write_bio(struct extent_page_data *epd)
4163 {
4164         if (epd->bio) {
4165                 int rw = WRITE;
4166                 int ret;
4167
4168                 if (epd->sync_io)
4169                         rw = WRITE_SYNC;
4170
4171                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
4172                 BUG_ON(ret < 0); /* -ENOMEM */
4173                 epd->bio = NULL;
4174         }
4175 }
4176
4177 static noinline void flush_write_bio(void *data)
4178 {
4179         struct extent_page_data *epd = data;
4180         flush_epd_write_bio(epd);
4181 }
4182
4183 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4184                           get_extent_t *get_extent,
4185                           struct writeback_control *wbc)
4186 {
4187         int ret;
4188         struct extent_page_data epd = {
4189                 .bio = NULL,
4190                 .tree = tree,
4191                 .get_extent = get_extent,
4192                 .extent_locked = 0,
4193                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4194                 .bio_flags = 0,
4195         };
4196
4197         ret = __extent_writepage(page, wbc, &epd);
4198
4199         flush_epd_write_bio(&epd);
4200         return ret;
4201 }
4202
4203 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4204                               u64 start, u64 end, get_extent_t *get_extent,
4205                               int mode)
4206 {
4207         int ret = 0;
4208         struct address_space *mapping = inode->i_mapping;
4209         struct page *page;
4210         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
4211                 PAGE_CACHE_SHIFT;
4212
4213         struct extent_page_data epd = {
4214                 .bio = NULL,
4215                 .tree = tree,
4216                 .get_extent = get_extent,
4217                 .extent_locked = 1,
4218                 .sync_io = mode == WB_SYNC_ALL,
4219                 .bio_flags = 0,
4220         };
4221         struct writeback_control wbc_writepages = {
4222                 .sync_mode      = mode,
4223                 .nr_to_write    = nr_pages * 2,
4224                 .range_start    = start,
4225                 .range_end      = end + 1,
4226         };
4227
4228         while (start <= end) {
4229                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
4230                 if (clear_page_dirty_for_io(page))
4231                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4232                 else {
4233                         if (tree->ops && tree->ops->writepage_end_io_hook)
4234                                 tree->ops->writepage_end_io_hook(page, start,
4235                                                  start + PAGE_CACHE_SIZE - 1,
4236                                                  NULL, 1);
4237                         unlock_page(page);
4238                 }
4239                 page_cache_release(page);
4240                 start += PAGE_CACHE_SIZE;
4241         }
4242
4243         flush_epd_write_bio(&epd);
4244         return ret;
4245 }
4246
4247 int extent_writepages(struct extent_io_tree *tree,
4248                       struct address_space *mapping,
4249                       get_extent_t *get_extent,
4250                       struct writeback_control *wbc)
4251 {
4252         int ret = 0;
4253         struct extent_page_data epd = {
4254                 .bio = NULL,
4255                 .tree = tree,
4256                 .get_extent = get_extent,
4257                 .extent_locked = 0,
4258                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4259                 .bio_flags = 0,
4260         };
4261
4262         ret = extent_write_cache_pages(tree, mapping, wbc,
4263                                        __extent_writepage, &epd,
4264                                        flush_write_bio);
4265         flush_epd_write_bio(&epd);
4266         return ret;
4267 }
4268
4269 int extent_readpages(struct extent_io_tree *tree,
4270                      struct address_space *mapping,
4271                      struct list_head *pages, unsigned nr_pages,
4272                      get_extent_t get_extent)
4273 {
4274         struct bio *bio = NULL;
4275         unsigned page_idx;
4276         unsigned long bio_flags = 0;
4277         struct page *pagepool[16];
4278         struct page *page;
4279         struct extent_map *em_cached = NULL;
4280         int nr = 0;
4281         u64 prev_em_start = (u64)-1;
4282
4283         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4284                 page = list_entry(pages->prev, struct page, lru);
4285
4286                 prefetchw(&page->flags);
4287                 list_del(&page->lru);
4288                 if (add_to_page_cache_lru(page, mapping,
4289                                         page->index, GFP_NOFS)) {
4290                         page_cache_release(page);
4291                         continue;
4292                 }
4293
4294                 pagepool[nr++] = page;
4295                 if (nr < ARRAY_SIZE(pagepool))
4296                         continue;
4297                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4298                                    &bio, 0, &bio_flags, READ, &prev_em_start);
4299                 nr = 0;
4300         }
4301         if (nr)
4302                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4303                                    &bio, 0, &bio_flags, READ, &prev_em_start);
4304
4305         if (em_cached)
4306                 free_extent_map(em_cached);
4307
4308         BUG_ON(!list_empty(pages));
4309         if (bio)
4310                 return submit_one_bio(READ, bio, 0, bio_flags);
4311         return 0;
4312 }
4313
4314 /*
4315  * basic invalidatepage code, this waits on any locked or writeback
4316  * ranges corresponding to the page, and then deletes any extent state
4317  * records from the tree
4318  */
4319 int extent_invalidatepage(struct extent_io_tree *tree,
4320                           struct page *page, unsigned long offset)
4321 {
4322         struct extent_state *cached_state = NULL;
4323         u64 start = page_offset(page);
4324         u64 end = start + PAGE_CACHE_SIZE - 1;
4325         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4326
4327         start += ALIGN(offset, blocksize);
4328         if (start > end)
4329                 return 0;
4330
4331         lock_extent_bits(tree, start, end, 0, &cached_state);
4332         wait_on_page_writeback(page);
4333         clear_extent_bit(tree, start, end,
4334                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4335                          EXTENT_DO_ACCOUNTING,
4336                          1, 1, &cached_state, GFP_NOFS);
4337         return 0;
4338 }
4339
4340 /*
4341  * a helper for releasepage, this tests for areas of the page that
4342  * are locked or under IO and drops the related state bits if it is safe
4343  * to drop the page.
4344  */
4345 static int try_release_extent_state(struct extent_map_tree *map,
4346                                     struct extent_io_tree *tree,
4347                                     struct page *page, gfp_t mask)
4348 {
4349         u64 start = page_offset(page);
4350         u64 end = start + PAGE_CACHE_SIZE - 1;
4351         int ret = 1;
4352
4353         if (test_range_bit(tree, start, end,
4354                            EXTENT_IOBITS, 0, NULL))
4355                 ret = 0;
4356         else {
4357                 if ((mask & GFP_NOFS) == GFP_NOFS)
4358                         mask = GFP_NOFS;
4359                 /*
4360                  * at this point we can safely clear everything except the
4361                  * locked bit and the nodatasum bit
4362                  */
4363                 ret = clear_extent_bit(tree, start, end,
4364                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4365                                  0, 0, NULL, mask);
4366
4367                 /* if clear_extent_bit failed for enomem reasons,
4368                  * we can't allow the release to continue.
4369                  */
4370                 if (ret < 0)
4371                         ret = 0;
4372                 else
4373                         ret = 1;
4374         }
4375         return ret;
4376 }
4377
4378 /*
4379  * a helper for releasepage.  As long as there are no locked extents
4380  * in the range corresponding to the page, both state records and extent
4381  * map records are removed
4382  */
4383 int try_release_extent_mapping(struct extent_map_tree *map,
4384                                struct extent_io_tree *tree, struct page *page,
4385                                gfp_t mask)
4386 {
4387         struct extent_map *em;
4388         u64 start = page_offset(page);
4389         u64 end = start + PAGE_CACHE_SIZE - 1;
4390
4391         if (gfpflags_allow_blocking(mask) &&
4392             page->mapping->host->i_size > 16 * 1024 * 1024) {
4393                 u64 len;
4394                 while (start <= end) {
4395                         len = end - start + 1;
4396                         write_lock(&map->lock);
4397                         em = lookup_extent_mapping(map, start, len);
4398                         if (!em) {
4399                                 write_unlock(&map->lock);
4400                                 break;
4401                         }
4402                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4403                             em->start != start) {
4404                                 write_unlock(&map->lock);
4405                                 free_extent_map(em);
4406                                 break;
4407                         }
4408                         if (!test_range_bit(tree, em->start,
4409                                             extent_map_end(em) - 1,
4410                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4411                                             0, NULL)) {
4412                                 remove_extent_mapping(map, em);
4413                                 /* once for the rb tree */
4414                                 free_extent_map(em);
4415                         }
4416                         start = extent_map_end(em);
4417                         write_unlock(&map->lock);
4418
4419                         /* once for us */
4420                         free_extent_map(em);
4421                 }
4422         }
4423         return try_release_extent_state(map, tree, page, mask);
4424 }
4425
4426 /*
4427  * helper function for fiemap, which doesn't want to see any holes.
4428  * This maps until we find something past 'last'
4429  */
4430 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4431                                                 u64 offset,
4432                                                 u64 last,
4433                                                 get_extent_t *get_extent)
4434 {
4435         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4436         struct extent_map *em;
4437         u64 len;
4438
4439         if (offset >= last)
4440                 return NULL;
4441
4442         while (1) {
4443                 len = last - offset;
4444                 if (len == 0)
4445                         break;
4446                 len = ALIGN(len, sectorsize);
4447                 em = get_extent(inode, NULL, 0, offset, len, 0);
4448                 if (IS_ERR_OR_NULL(em))
4449                         return em;
4450
4451                 /* if this isn't a hole return it */
4452                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4453                     em->block_start != EXTENT_MAP_HOLE) {
4454                         return em;
4455                 }
4456
4457                 /* this is a hole, advance to the next extent */
4458                 offset = extent_map_end(em);
4459                 free_extent_map(em);
4460                 if (offset >= last)
4461                         break;
4462         }
4463         return NULL;
4464 }
4465
4466 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4467                 __u64 start, __u64 len, get_extent_t *get_extent)
4468 {
4469         int ret = 0;
4470         u64 off = start;
4471         u64 max = start + len;
4472         u32 flags = 0;
4473         u32 found_type;
4474         u64 last;
4475         u64 last_for_get_extent = 0;
4476         u64 disko = 0;
4477         u64 isize = i_size_read(inode);
4478         struct btrfs_key found_key;
4479         struct extent_map *em = NULL;
4480         struct extent_state *cached_state = NULL;
4481         struct btrfs_path *path;
4482         struct btrfs_root *root = BTRFS_I(inode)->root;
4483         int end = 0;
4484         u64 em_start = 0;
4485         u64 em_len = 0;
4486         u64 em_end = 0;
4487
4488         if (len == 0)
4489                 return -EINVAL;
4490
4491         path = btrfs_alloc_path();
4492         if (!path)
4493                 return -ENOMEM;
4494         path->leave_spinning = 1;
4495
4496         start = round_down(start, BTRFS_I(inode)->root->sectorsize);
4497         len = round_up(max, BTRFS_I(inode)->root->sectorsize) - start;
4498
4499         /*
4500          * lookup the last file extent.  We're not using i_size here
4501          * because there might be preallocation past i_size
4502          */
4503         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode), -1,
4504                                        0);
4505         if (ret < 0) {
4506                 btrfs_free_path(path);
4507                 return ret;
4508         }
4509         WARN_ON(!ret);
4510         path->slots[0]--;
4511         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4512         found_type = found_key.type;
4513
4514         /* No extents, but there might be delalloc bits */
4515         if (found_key.objectid != btrfs_ino(inode) ||
4516             found_type != BTRFS_EXTENT_DATA_KEY) {
4517                 /* have to trust i_size as the end */
4518                 last = (u64)-1;
4519                 last_for_get_extent = isize;
4520         } else {
4521                 /*
4522                  * remember the start of the last extent.  There are a
4523                  * bunch of different factors that go into the length of the
4524                  * extent, so its much less complex to remember where it started
4525                  */
4526                 last = found_key.offset;
4527                 last_for_get_extent = last + 1;
4528         }
4529         btrfs_release_path(path);
4530
4531         /*
4532          * we might have some extents allocated but more delalloc past those
4533          * extents.  so, we trust isize unless the start of the last extent is
4534          * beyond isize
4535          */
4536         if (last < isize) {
4537                 last = (u64)-1;
4538                 last_for_get_extent = isize;
4539         }
4540
4541         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4542                          &cached_state);
4543
4544         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4545                                    get_extent);
4546         if (!em)
4547                 goto out;
4548         if (IS_ERR(em)) {
4549                 ret = PTR_ERR(em);
4550                 goto out;
4551         }
4552
4553         while (!end) {
4554                 u64 offset_in_extent = 0;
4555
4556                 /* break if the extent we found is outside the range */
4557                 if (em->start >= max || extent_map_end(em) < off)
4558                         break;
4559
4560                 /*
4561                  * get_extent may return an extent that starts before our
4562                  * requested range.  We have to make sure the ranges
4563                  * we return to fiemap always move forward and don't
4564                  * overlap, so adjust the offsets here
4565                  */
4566                 em_start = max(em->start, off);
4567
4568                 /*
4569                  * record the offset from the start of the extent
4570                  * for adjusting the disk offset below.  Only do this if the
4571                  * extent isn't compressed since our in ram offset may be past
4572                  * what we have actually allocated on disk.
4573                  */
4574                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4575                         offset_in_extent = em_start - em->start;
4576                 em_end = extent_map_end(em);
4577                 em_len = em_end - em_start;
4578                 disko = 0;
4579                 flags = 0;
4580
4581                 /*
4582                  * bump off for our next call to get_extent
4583                  */
4584                 off = extent_map_end(em);
4585                 if (off >= max)
4586                         end = 1;
4587
4588                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4589                         end = 1;
4590                         flags |= FIEMAP_EXTENT_LAST;
4591                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4592                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4593                                   FIEMAP_EXTENT_NOT_ALIGNED);
4594                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4595                         flags |= (FIEMAP_EXTENT_DELALLOC |
4596                                   FIEMAP_EXTENT_UNKNOWN);
4597                 } else if (fieinfo->fi_extents_max) {
4598                         u64 bytenr = em->block_start -
4599                                 (em->start - em->orig_start);
4600
4601                         disko = em->block_start + offset_in_extent;
4602
4603                         /*
4604                          * As btrfs supports shared space, this information
4605                          * can be exported to userspace tools via
4606                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4607                          * then we're just getting a count and we can skip the
4608                          * lookup stuff.
4609                          */
4610                         ret = btrfs_check_shared(NULL, root->fs_info,
4611                                                  root->objectid,
4612                                                  btrfs_ino(inode), bytenr);
4613                         if (ret < 0)
4614                                 goto out_free;
4615                         if (ret)
4616                                 flags |= FIEMAP_EXTENT_SHARED;
4617                         ret = 0;
4618                 }
4619                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4620                         flags |= FIEMAP_EXTENT_ENCODED;
4621                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4622                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4623
4624                 free_extent_map(em);
4625                 em = NULL;
4626                 if ((em_start >= last) || em_len == (u64)-1 ||
4627                    (last == (u64)-1 && isize <= em_end)) {
4628                         flags |= FIEMAP_EXTENT_LAST;
4629                         end = 1;
4630                 }
4631
4632                 /* now scan forward to see if this is really the last extent. */
4633                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4634                                            get_extent);
4635                 if (IS_ERR(em)) {
4636                         ret = PTR_ERR(em);
4637                         goto out;
4638                 }
4639                 if (!em) {
4640                         flags |= FIEMAP_EXTENT_LAST;
4641                         end = 1;
4642                 }
4643                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4644                                               em_len, flags);
4645                 if (ret) {
4646                         if (ret == 1)
4647                                 ret = 0;
4648                         goto out_free;
4649                 }
4650         }
4651 out_free:
4652         free_extent_map(em);
4653 out:
4654         btrfs_free_path(path);
4655         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4656                              &cached_state, GFP_NOFS);
4657         return ret;
4658 }
4659
4660 static void __free_extent_buffer(struct extent_buffer *eb)
4661 {
4662         btrfs_leak_debug_del(&eb->leak_list);
4663         kmem_cache_free(extent_buffer_cache, eb);
4664 }
4665
4666 int extent_buffer_under_io(struct extent_buffer *eb)
4667 {
4668         return (atomic_read(&eb->io_pages) ||
4669                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4670                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4671 }
4672
4673 /*
4674  * Helper for releasing extent buffer page.
4675  */
4676 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4677 {
4678         unsigned long index;
4679         struct page *page;
4680         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4681
4682         BUG_ON(extent_buffer_under_io(eb));
4683
4684         index = num_extent_pages(eb->start, eb->len);
4685         if (index == 0)
4686                 return;
4687
4688         do {
4689                 index--;
4690                 page = eb->pages[index];
4691                 if (!page)
4692                         continue;
4693                 if (mapped)
4694                         spin_lock(&page->mapping->private_lock);
4695                 /*
4696                  * We do this since we'll remove the pages after we've
4697                  * removed the eb from the radix tree, so we could race
4698                  * and have this page now attached to the new eb.  So
4699                  * only clear page_private if it's still connected to
4700                  * this eb.
4701                  */
4702                 if (PagePrivate(page) &&
4703                     page->private == (unsigned long)eb) {
4704                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4705                         BUG_ON(PageDirty(page));
4706                         BUG_ON(PageWriteback(page));
4707                         /*
4708                          * We need to make sure we haven't be attached
4709                          * to a new eb.
4710                          */
4711                         ClearPagePrivate(page);
4712                         set_page_private(page, 0);
4713                         /* One for the page private */
4714                         page_cache_release(page);
4715                 }
4716
4717                 if (mapped)
4718                         spin_unlock(&page->mapping->private_lock);
4719
4720                 /* One for when we alloced the page */
4721                 page_cache_release(page);
4722         } while (index != 0);
4723 }
4724
4725 /*
4726  * Helper for releasing the extent buffer.
4727  */
4728 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4729 {
4730         btrfs_release_extent_buffer_page(eb);
4731         __free_extent_buffer(eb);
4732 }
4733
4734 static struct extent_buffer *
4735 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4736                       unsigned long len)
4737 {
4738         struct extent_buffer *eb = NULL;
4739
4740         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4741         eb->start = start;
4742         eb->len = len;
4743         eb->fs_info = fs_info;
4744         eb->bflags = 0;
4745         rwlock_init(&eb->lock);
4746         atomic_set(&eb->write_locks, 0);
4747         atomic_set(&eb->read_locks, 0);
4748         atomic_set(&eb->blocking_readers, 0);
4749         atomic_set(&eb->blocking_writers, 0);
4750         atomic_set(&eb->spinning_readers, 0);
4751         atomic_set(&eb->spinning_writers, 0);
4752         eb->lock_nested = 0;
4753         init_waitqueue_head(&eb->write_lock_wq);
4754         init_waitqueue_head(&eb->read_lock_wq);
4755
4756         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4757
4758         spin_lock_init(&eb->refs_lock);
4759         atomic_set(&eb->refs, 1);
4760         atomic_set(&eb->io_pages, 0);
4761
4762         /*
4763          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4764          */
4765         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4766                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4767         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4768
4769         return eb;
4770 }
4771
4772 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4773 {
4774         unsigned long i;
4775         struct page *p;
4776         struct extent_buffer *new;
4777         unsigned long num_pages = num_extent_pages(src->start, src->len);
4778
4779         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4780         if (new == NULL)
4781                 return NULL;
4782
4783         for (i = 0; i < num_pages; i++) {
4784                 p = alloc_page(GFP_NOFS);
4785                 if (!p) {
4786                         btrfs_release_extent_buffer(new);
4787                         return NULL;
4788                 }
4789                 attach_extent_buffer_page(new, p);
4790                 WARN_ON(PageDirty(p));
4791                 SetPageUptodate(p);
4792                 new->pages[i] = p;
4793         }
4794
4795         copy_extent_buffer(new, src, 0, 0, src->len);
4796         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4797         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4798
4799         return new;
4800 }
4801
4802 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4803                                                 u64 start)
4804 {
4805         struct extent_buffer *eb;
4806         unsigned long len;
4807         unsigned long num_pages;
4808         unsigned long i;
4809
4810         if (!fs_info) {
4811                 /*
4812                  * Called only from tests that don't always have a fs_info
4813                  * available, but we know that nodesize is 4096
4814                  */
4815                 len = 4096;
4816         } else {
4817                 len = fs_info->tree_root->nodesize;
4818         }
4819         num_pages = num_extent_pages(0, len);
4820
4821         eb = __alloc_extent_buffer(fs_info, start, len);
4822         if (!eb)
4823                 return NULL;
4824
4825         for (i = 0; i < num_pages; i++) {
4826                 eb->pages[i] = alloc_page(GFP_NOFS);
4827                 if (!eb->pages[i])
4828                         goto err;
4829         }
4830         set_extent_buffer_uptodate(eb);
4831         btrfs_set_header_nritems(eb, 0);
4832         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4833
4834         return eb;
4835 err:
4836         for (; i > 0; i--)
4837                 __free_page(eb->pages[i - 1]);
4838         __free_extent_buffer(eb);
4839         return NULL;
4840 }
4841
4842 static void check_buffer_tree_ref(struct extent_buffer *eb)
4843 {
4844         int refs;
4845         /* the ref bit is tricky.  We have to make sure it is set
4846          * if we have the buffer dirty.   Otherwise the
4847          * code to free a buffer can end up dropping a dirty
4848          * page
4849          *
4850          * Once the ref bit is set, it won't go away while the
4851          * buffer is dirty or in writeback, and it also won't
4852          * go away while we have the reference count on the
4853          * eb bumped.
4854          *
4855          * We can't just set the ref bit without bumping the
4856          * ref on the eb because free_extent_buffer might
4857          * see the ref bit and try to clear it.  If this happens
4858          * free_extent_buffer might end up dropping our original
4859          * ref by mistake and freeing the page before we are able
4860          * to add one more ref.
4861          *
4862          * So bump the ref count first, then set the bit.  If someone
4863          * beat us to it, drop the ref we added.
4864          */
4865         refs = atomic_read(&eb->refs);
4866         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4867                 return;
4868
4869         spin_lock(&eb->refs_lock);
4870         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4871                 atomic_inc(&eb->refs);
4872         spin_unlock(&eb->refs_lock);
4873 }
4874
4875 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4876                 struct page *accessed)
4877 {
4878         unsigned long num_pages, i;
4879
4880         check_buffer_tree_ref(eb);
4881
4882         num_pages = num_extent_pages(eb->start, eb->len);
4883         for (i = 0; i < num_pages; i++) {
4884                 struct page *p = eb->pages[i];
4885
4886                 if (p != accessed)
4887                         mark_page_accessed(p);
4888         }
4889 }
4890
4891 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4892                                          u64 start)
4893 {
4894         struct extent_buffer *eb;
4895
4896         rcu_read_lock();
4897         eb = radix_tree_lookup(&fs_info->buffer_radix,
4898                                start >> PAGE_CACHE_SHIFT);
4899         if (eb && atomic_inc_not_zero(&eb->refs)) {
4900                 rcu_read_unlock();
4901                 /*
4902                  * Lock our eb's refs_lock to avoid races with
4903                  * free_extent_buffer. When we get our eb it might be flagged
4904                  * with EXTENT_BUFFER_STALE and another task running
4905                  * free_extent_buffer might have seen that flag set,
4906                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4907                  * writeback flags not set) and it's still in the tree (flag
4908                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4909                  * of decrementing the extent buffer's reference count twice.
4910                  * So here we could race and increment the eb's reference count,
4911                  * clear its stale flag, mark it as dirty and drop our reference
4912                  * before the other task finishes executing free_extent_buffer,
4913                  * which would later result in an attempt to free an extent
4914                  * buffer that is dirty.
4915                  */
4916                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4917                         spin_lock(&eb->refs_lock);
4918                         spin_unlock(&eb->refs_lock);
4919                 }
4920                 mark_extent_buffer_accessed(eb, NULL);
4921                 return eb;
4922         }
4923         rcu_read_unlock();
4924
4925         return NULL;
4926 }
4927
4928 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4929 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4930                                                u64 start)
4931 {
4932         struct extent_buffer *eb, *exists = NULL;
4933         int ret;
4934
4935         eb = find_extent_buffer(fs_info, start);
4936         if (eb)
4937                 return eb;
4938         eb = alloc_dummy_extent_buffer(fs_info, start);
4939         if (!eb)
4940                 return ERR_PTR(-ENOMEM);
4941         eb->fs_info = fs_info;
4942 again:
4943         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4944         if (ret) {
4945                 exists = ERR_PTR(ret);
4946                 goto free_eb;
4947         }
4948         spin_lock(&fs_info->buffer_lock);
4949         ret = radix_tree_insert(&fs_info->buffer_radix,
4950                                 start >> PAGE_CACHE_SHIFT, eb);
4951         spin_unlock(&fs_info->buffer_lock);
4952         radix_tree_preload_end();
4953         if (ret == -EEXIST) {
4954                 exists = find_extent_buffer(fs_info, start);
4955                 if (exists)
4956                         goto free_eb;
4957                 else
4958                         goto again;
4959         }
4960         check_buffer_tree_ref(eb);
4961         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4962
4963         /*
4964          * We will free dummy extent buffer's if they come into
4965          * free_extent_buffer with a ref count of 2, but if we are using this we
4966          * want the buffers to stay in memory until we're done with them, so
4967          * bump the ref count again.
4968          */
4969         atomic_inc(&eb->refs);
4970         return eb;
4971 free_eb:
4972         btrfs_release_extent_buffer(eb);
4973         return exists;
4974 }
4975 #endif
4976
4977 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4978                                           u64 start)
4979 {
4980         unsigned long len = fs_info->tree_root->nodesize;
4981         unsigned long num_pages = num_extent_pages(start, len);
4982         unsigned long i;
4983         unsigned long index = start >> PAGE_CACHE_SHIFT;
4984         struct extent_buffer *eb;
4985         struct extent_buffer *exists = NULL;
4986         struct page *p;
4987         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4988         int uptodate = 1;
4989         int ret;
4990
4991         eb = find_extent_buffer(fs_info, start);
4992         if (eb)
4993                 return eb;
4994
4995         eb = __alloc_extent_buffer(fs_info, start, len);
4996         if (!eb)
4997                 return NULL;
4998
4999         for (i = 0; i < num_pages; i++, index++) {
5000                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
5001                 if (!p)
5002                         goto free_eb;
5003
5004                 spin_lock(&mapping->private_lock);
5005                 if (PagePrivate(p)) {
5006                         /*
5007                          * We could have already allocated an eb for this page
5008                          * and attached one so lets see if we can get a ref on
5009                          * the existing eb, and if we can we know it's good and
5010                          * we can just return that one, else we know we can just
5011                          * overwrite page->private.
5012                          */
5013                         exists = (struct extent_buffer *)p->private;
5014                         if (atomic_inc_not_zero(&exists->refs)) {
5015                                 spin_unlock(&mapping->private_lock);
5016                                 unlock_page(p);
5017                                 page_cache_release(p);
5018                                 mark_extent_buffer_accessed(exists, p);
5019                                 goto free_eb;
5020                         }
5021                         exists = NULL;
5022
5023                         /*
5024                          * Do this so attach doesn't complain and we need to
5025                          * drop the ref the old guy had.
5026                          */
5027                         ClearPagePrivate(p);
5028                         WARN_ON(PageDirty(p));
5029                         page_cache_release(p);
5030                 }
5031                 attach_extent_buffer_page(eb, p);
5032                 spin_unlock(&mapping->private_lock);
5033                 WARN_ON(PageDirty(p));
5034                 eb->pages[i] = p;
5035                 if (!PageUptodate(p))
5036                         uptodate = 0;
5037
5038                 /*
5039                  * see below about how we avoid a nasty race with release page
5040                  * and why we unlock later
5041                  */
5042         }
5043         if (uptodate)
5044                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5045 again:
5046         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
5047         if (ret)
5048                 goto free_eb;
5049
5050         spin_lock(&fs_info->buffer_lock);
5051         ret = radix_tree_insert(&fs_info->buffer_radix,
5052                                 start >> PAGE_CACHE_SHIFT, eb);
5053         spin_unlock(&fs_info->buffer_lock);
5054         radix_tree_preload_end();
5055         if (ret == -EEXIST) {
5056                 exists = find_extent_buffer(fs_info, start);
5057                 if (exists)
5058                         goto free_eb;
5059                 else
5060                         goto again;
5061         }
5062         /* add one reference for the tree */
5063         check_buffer_tree_ref(eb);
5064         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5065
5066         /*
5067          * there is a race where release page may have
5068          * tried to find this extent buffer in the radix
5069          * but failed.  It will tell the VM it is safe to
5070          * reclaim the, and it will clear the page private bit.
5071          * We must make sure to set the page private bit properly
5072          * after the extent buffer is in the radix tree so
5073          * it doesn't get lost
5074          */
5075         SetPageChecked(eb->pages[0]);
5076         for (i = 1; i < num_pages; i++) {
5077                 p = eb->pages[i];
5078                 ClearPageChecked(p);
5079                 unlock_page(p);
5080         }
5081         unlock_page(eb->pages[0]);
5082         return eb;
5083
5084 free_eb:
5085         WARN_ON(!atomic_dec_and_test(&eb->refs));
5086         for (i = 0; i < num_pages; i++) {
5087                 if (eb->pages[i])
5088                         unlock_page(eb->pages[i]);
5089         }
5090
5091         btrfs_release_extent_buffer(eb);
5092         return exists;
5093 }
5094
5095 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5096 {
5097         struct extent_buffer *eb =
5098                         container_of(head, struct extent_buffer, rcu_head);
5099
5100         __free_extent_buffer(eb);
5101 }
5102
5103 /* Expects to have eb->eb_lock already held */
5104 static int release_extent_buffer(struct extent_buffer *eb)
5105 {
5106         WARN_ON(atomic_read(&eb->refs) == 0);
5107         if (atomic_dec_and_test(&eb->refs)) {
5108                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5109                         struct btrfs_fs_info *fs_info = eb->fs_info;
5110
5111                         spin_unlock(&eb->refs_lock);
5112
5113                         spin_lock(&fs_info->buffer_lock);
5114                         radix_tree_delete(&fs_info->buffer_radix,
5115                                           eb->start >> PAGE_CACHE_SHIFT);
5116                         spin_unlock(&fs_info->buffer_lock);
5117                 } else {
5118                         spin_unlock(&eb->refs_lock);
5119                 }
5120
5121                 /* Should be safe to release our pages at this point */
5122                 btrfs_release_extent_buffer_page(eb);
5123 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5124                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5125                         __free_extent_buffer(eb);
5126                         return 1;
5127                 }
5128 #endif
5129                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5130                 return 1;
5131         }
5132         spin_unlock(&eb->refs_lock);
5133
5134         return 0;
5135 }
5136
5137 void free_extent_buffer(struct extent_buffer *eb)
5138 {
5139         int refs;
5140         int old;
5141         if (!eb)
5142                 return;
5143
5144         while (1) {
5145                 refs = atomic_read(&eb->refs);
5146                 if (refs <= 3)
5147                         break;
5148                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5149                 if (old == refs)
5150                         return;
5151         }
5152
5153         spin_lock(&eb->refs_lock);
5154         if (atomic_read(&eb->refs) == 2 &&
5155             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5156                 atomic_dec(&eb->refs);
5157
5158         if (atomic_read(&eb->refs) == 2 &&
5159             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5160             !extent_buffer_under_io(eb) &&
5161             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5162                 atomic_dec(&eb->refs);
5163
5164         /*
5165          * I know this is terrible, but it's temporary until we stop tracking
5166          * the uptodate bits and such for the extent buffers.
5167          */
5168         release_extent_buffer(eb);
5169 }
5170
5171 void free_extent_buffer_stale(struct extent_buffer *eb)
5172 {
5173         if (!eb)
5174                 return;
5175
5176         spin_lock(&eb->refs_lock);
5177         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5178
5179         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5180             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5181                 atomic_dec(&eb->refs);
5182         release_extent_buffer(eb);
5183 }
5184
5185 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5186 {
5187         unsigned long i;
5188         unsigned long num_pages;
5189         struct page *page;
5190
5191         num_pages = num_extent_pages(eb->start, eb->len);
5192
5193         for (i = 0; i < num_pages; i++) {
5194                 page = eb->pages[i];
5195                 if (!PageDirty(page))
5196                         continue;
5197
5198                 lock_page(page);
5199                 WARN_ON(!PagePrivate(page));
5200
5201                 clear_page_dirty_for_io(page);
5202                 spin_lock_irq(&page->mapping->tree_lock);
5203                 if (!PageDirty(page)) {
5204                         radix_tree_tag_clear(&page->mapping->page_tree,
5205                                                 page_index(page),
5206                                                 PAGECACHE_TAG_DIRTY);
5207                 }
5208                 spin_unlock_irq(&page->mapping->tree_lock);
5209                 ClearPageError(page);
5210                 unlock_page(page);
5211         }
5212         WARN_ON(atomic_read(&eb->refs) == 0);
5213 }
5214
5215 int set_extent_buffer_dirty(struct extent_buffer *eb)
5216 {
5217         unsigned long i;
5218         unsigned long num_pages;
5219         int was_dirty = 0;
5220
5221         check_buffer_tree_ref(eb);
5222
5223         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5224
5225         num_pages = num_extent_pages(eb->start, eb->len);
5226         WARN_ON(atomic_read(&eb->refs) == 0);
5227         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5228
5229         for (i = 0; i < num_pages; i++)
5230                 set_page_dirty(eb->pages[i]);
5231         return was_dirty;
5232 }
5233
5234 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
5235 {
5236         unsigned long i;
5237         struct page *page;
5238         unsigned long num_pages;
5239
5240         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5241         num_pages = num_extent_pages(eb->start, eb->len);
5242         for (i = 0; i < num_pages; i++) {
5243                 page = eb->pages[i];
5244                 if (page)
5245                         ClearPageUptodate(page);
5246         }
5247         return 0;
5248 }
5249
5250 int set_extent_buffer_uptodate(struct extent_buffer *eb)
5251 {
5252         unsigned long i;
5253         struct page *page;
5254         unsigned long num_pages;
5255
5256         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5257         num_pages = num_extent_pages(eb->start, eb->len);
5258         for (i = 0; i < num_pages; i++) {
5259                 page = eb->pages[i];
5260                 SetPageUptodate(page);
5261         }
5262         return 0;
5263 }
5264
5265 int extent_buffer_uptodate(struct extent_buffer *eb)
5266 {
5267         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5268 }
5269
5270 int read_extent_buffer_pages(struct extent_io_tree *tree,
5271                              struct extent_buffer *eb, u64 start, int wait,
5272                              get_extent_t *get_extent, int mirror_num)
5273 {
5274         unsigned long i;
5275         unsigned long start_i;
5276         struct page *page;
5277         int err;
5278         int ret = 0;
5279         int locked_pages = 0;
5280         int all_uptodate = 1;
5281         unsigned long num_pages;
5282         unsigned long num_reads = 0;
5283         struct bio *bio = NULL;
5284         unsigned long bio_flags = 0;
5285
5286         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5287                 return 0;
5288
5289         if (start) {
5290                 WARN_ON(start < eb->start);
5291                 start_i = (start >> PAGE_CACHE_SHIFT) -
5292                         (eb->start >> PAGE_CACHE_SHIFT);
5293         } else {
5294                 start_i = 0;
5295         }
5296
5297         num_pages = num_extent_pages(eb->start, eb->len);
5298         for (i = start_i; i < num_pages; i++) {
5299                 page = eb->pages[i];
5300                 if (wait == WAIT_NONE) {
5301                         if (!trylock_page(page))
5302                                 goto unlock_exit;
5303                 } else {
5304                         lock_page(page);
5305                 }
5306                 locked_pages++;
5307         }
5308         /*
5309          * We need to firstly lock all pages to make sure that
5310          * the uptodate bit of our pages won't be affected by
5311          * clear_extent_buffer_uptodate().
5312          */
5313         for (i = start_i; i < num_pages; i++) {
5314                 page = eb->pages[i];
5315                 if (!PageUptodate(page)) {
5316                         num_reads++;
5317                         all_uptodate = 0;
5318                 }
5319         }
5320
5321         if (all_uptodate) {
5322                 if (start_i == 0)
5323                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5324                 goto unlock_exit;
5325         }
5326
5327         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5328         eb->read_mirror = 0;
5329         atomic_set(&eb->io_pages, num_reads);
5330         for (i = start_i; i < num_pages; i++) {
5331                 page = eb->pages[i];
5332                 if (!PageUptodate(page)) {
5333                         ClearPageError(page);
5334                         err = __extent_read_full_page(tree, page,
5335                                                       get_extent, &bio,
5336                                                       mirror_num, &bio_flags,
5337                                                       READ | REQ_META);
5338                         if (err)
5339                                 ret = err;
5340                 } else {
5341                         unlock_page(page);
5342                 }
5343         }
5344
5345         if (bio) {
5346                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
5347                                      bio_flags);
5348                 if (err)
5349                         return err;
5350         }
5351
5352         if (ret || wait != WAIT_COMPLETE)
5353                 return ret;
5354
5355         for (i = start_i; i < num_pages; i++) {
5356                 page = eb->pages[i];
5357                 wait_on_page_locked(page);
5358                 if (!PageUptodate(page))
5359                         ret = -EIO;
5360         }
5361
5362         return ret;
5363
5364 unlock_exit:
5365         i = start_i;
5366         while (locked_pages > 0) {
5367                 page = eb->pages[i];
5368                 i++;
5369                 unlock_page(page);
5370                 locked_pages--;
5371         }
5372         return ret;
5373 }
5374
5375 void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
5376                         unsigned long start, unsigned long len)
5377 {
5378         size_t cur;
5379         size_t offset;
5380         struct page *page;
5381         char *kaddr;
5382         char *dst = (char *)dstv;
5383         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5384         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5385
5386         WARN_ON(start > eb->len);
5387         WARN_ON(start + len > eb->start + eb->len);
5388
5389         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5390
5391         while (len > 0) {
5392                 page = eb->pages[i];
5393
5394                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5395                 kaddr = page_address(page);
5396                 memcpy(dst, kaddr + offset, cur);
5397
5398                 dst += cur;
5399                 len -= cur;
5400                 offset = 0;
5401                 i++;
5402         }
5403 }
5404
5405 int read_extent_buffer_to_user(const struct extent_buffer *eb,
5406                                void __user *dstv,
5407                                unsigned long start, unsigned long len)
5408 {
5409         size_t cur;
5410         size_t offset;
5411         struct page *page;
5412         char *kaddr;
5413         char __user *dst = (char __user *)dstv;
5414         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5415         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5416         int ret = 0;
5417
5418         WARN_ON(start > eb->len);
5419         WARN_ON(start + len > eb->start + eb->len);
5420
5421         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5422
5423         while (len > 0) {
5424                 page = eb->pages[i];
5425
5426                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5427                 kaddr = page_address(page);
5428                 if (copy_to_user(dst, kaddr + offset, cur)) {
5429                         ret = -EFAULT;
5430                         break;
5431                 }
5432
5433                 dst += cur;
5434                 len -= cur;
5435                 offset = 0;
5436                 i++;
5437         }
5438
5439         return ret;
5440 }
5441
5442 int map_private_extent_buffer(const struct extent_buffer *eb,
5443                               unsigned long start, unsigned long min_len,
5444                               char **map, unsigned long *map_start,
5445                               unsigned long *map_len)
5446 {
5447         size_t offset = start & (PAGE_CACHE_SIZE - 1);
5448         char *kaddr;
5449         struct page *p;
5450         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5451         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5452         unsigned long end_i = (start_offset + start + min_len - 1) >>
5453                 PAGE_CACHE_SHIFT;
5454
5455         if (i != end_i)
5456                 return -EINVAL;
5457
5458         if (i == 0) {
5459                 offset = start_offset;
5460                 *map_start = 0;
5461         } else {
5462                 offset = 0;
5463                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
5464         }
5465
5466         if (start + min_len > eb->len) {
5467                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
5468                        "wanted %lu %lu\n",
5469                        eb->start, eb->len, start, min_len);
5470                 return -EINVAL;
5471         }
5472
5473         p = eb->pages[i];
5474         kaddr = page_address(p);
5475         *map = kaddr + offset;
5476         *map_len = PAGE_CACHE_SIZE - offset;
5477         return 0;
5478 }
5479
5480 int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
5481                          unsigned long start, unsigned long len)
5482 {
5483         size_t cur;
5484         size_t offset;
5485         struct page *page;
5486         char *kaddr;
5487         char *ptr = (char *)ptrv;
5488         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5489         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5490         int ret = 0;
5491
5492         WARN_ON(start > eb->len);
5493         WARN_ON(start + len > eb->start + eb->len);
5494
5495         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5496
5497         while (len > 0) {
5498                 page = eb->pages[i];
5499
5500                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5501
5502                 kaddr = page_address(page);
5503                 ret = memcmp(ptr, kaddr + offset, cur);
5504                 if (ret)
5505                         break;
5506
5507                 ptr += cur;
5508                 len -= cur;
5509                 offset = 0;
5510                 i++;
5511         }
5512         return ret;
5513 }
5514
5515 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5516                          unsigned long start, unsigned long len)
5517 {
5518         size_t cur;
5519         size_t offset;
5520         struct page *page;
5521         char *kaddr;
5522         char *src = (char *)srcv;
5523         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5524         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5525
5526         WARN_ON(start > eb->len);
5527         WARN_ON(start + len > eb->start + eb->len);
5528
5529         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5530
5531         while (len > 0) {
5532                 page = eb->pages[i];
5533                 WARN_ON(!PageUptodate(page));
5534
5535                 cur = min(len, PAGE_CACHE_SIZE - offset);
5536                 kaddr = page_address(page);
5537                 memcpy(kaddr + offset, src, cur);
5538
5539                 src += cur;
5540                 len -= cur;
5541                 offset = 0;
5542                 i++;
5543         }
5544 }
5545
5546 void memset_extent_buffer(struct extent_buffer *eb, char c,
5547                           unsigned long start, unsigned long len)
5548 {
5549         size_t cur;
5550         size_t offset;
5551         struct page *page;
5552         char *kaddr;
5553         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5554         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5555
5556         WARN_ON(start > eb->len);
5557         WARN_ON(start + len > eb->start + eb->len);
5558
5559         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5560
5561         while (len > 0) {
5562                 page = eb->pages[i];
5563                 WARN_ON(!PageUptodate(page));
5564
5565                 cur = min(len, PAGE_CACHE_SIZE - offset);
5566                 kaddr = page_address(page);
5567                 memset(kaddr + offset, c, cur);
5568
5569                 len -= cur;
5570                 offset = 0;
5571                 i++;
5572         }
5573 }
5574
5575 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5576                         unsigned long dst_offset, unsigned long src_offset,
5577                         unsigned long len)
5578 {
5579         u64 dst_len = dst->len;
5580         size_t cur;
5581         size_t offset;
5582         struct page *page;
5583         char *kaddr;
5584         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5585         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5586
5587         WARN_ON(src->len != dst_len);
5588
5589         offset = (start_offset + dst_offset) &
5590                 (PAGE_CACHE_SIZE - 1);
5591
5592         while (len > 0) {
5593                 page = dst->pages[i];
5594                 WARN_ON(!PageUptodate(page));
5595
5596                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5597
5598                 kaddr = page_address(page);
5599                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5600
5601                 src_offset += cur;
5602                 len -= cur;
5603                 offset = 0;
5604                 i++;
5605         }
5606 }
5607
5608 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5609 {
5610         unsigned long distance = (src > dst) ? src - dst : dst - src;
5611         return distance < len;
5612 }
5613
5614 static void copy_pages(struct page *dst_page, struct page *src_page,
5615                        unsigned long dst_off, unsigned long src_off,
5616                        unsigned long len)
5617 {
5618         char *dst_kaddr = page_address(dst_page);
5619         char *src_kaddr;
5620         int must_memmove = 0;
5621
5622         if (dst_page != src_page) {
5623                 src_kaddr = page_address(src_page);
5624         } else {
5625                 src_kaddr = dst_kaddr;
5626                 if (areas_overlap(src_off, dst_off, len))
5627                         must_memmove = 1;
5628         }
5629
5630         if (must_memmove)
5631                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5632         else
5633                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5634 }
5635
5636 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5637                            unsigned long src_offset, unsigned long len)
5638 {
5639         size_t cur;
5640         size_t dst_off_in_page;
5641         size_t src_off_in_page;
5642         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5643         unsigned long dst_i;
5644         unsigned long src_i;
5645
5646         if (src_offset + len > dst->len) {
5647                 btrfs_err(dst->fs_info,
5648                         "memmove bogus src_offset %lu move "
5649                        "len %lu dst len %lu", src_offset, len, dst->len);
5650                 BUG_ON(1);
5651         }
5652         if (dst_offset + len > dst->len) {
5653                 btrfs_err(dst->fs_info,
5654                         "memmove bogus dst_offset %lu move "
5655                        "len %lu dst len %lu", dst_offset, len, dst->len);
5656                 BUG_ON(1);
5657         }
5658
5659         while (len > 0) {
5660                 dst_off_in_page = (start_offset + dst_offset) &
5661                         (PAGE_CACHE_SIZE - 1);
5662                 src_off_in_page = (start_offset + src_offset) &
5663                         (PAGE_CACHE_SIZE - 1);
5664
5665                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5666                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5667
5668                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5669                                                src_off_in_page));
5670                 cur = min_t(unsigned long, cur,
5671                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5672
5673                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5674                            dst_off_in_page, src_off_in_page, cur);
5675
5676                 src_offset += cur;
5677                 dst_offset += cur;
5678                 len -= cur;
5679         }
5680 }
5681
5682 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5683                            unsigned long src_offset, unsigned long len)
5684 {
5685         size_t cur;
5686         size_t dst_off_in_page;
5687         size_t src_off_in_page;
5688         unsigned long dst_end = dst_offset + len - 1;
5689         unsigned long src_end = src_offset + len - 1;
5690         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5691         unsigned long dst_i;
5692         unsigned long src_i;
5693
5694         if (src_offset + len > dst->len) {
5695                 btrfs_err(dst->fs_info, "memmove bogus src_offset %lu move "
5696                        "len %lu len %lu", src_offset, len, dst->len);
5697                 BUG_ON(1);
5698         }
5699         if (dst_offset + len > dst->len) {
5700                 btrfs_err(dst->fs_info, "memmove bogus dst_offset %lu move "
5701                        "len %lu len %lu", dst_offset, len, dst->len);
5702                 BUG_ON(1);
5703         }
5704         if (dst_offset < src_offset) {
5705                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5706                 return;
5707         }
5708         while (len > 0) {
5709                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5710                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5711
5712                 dst_off_in_page = (start_offset + dst_end) &
5713                         (PAGE_CACHE_SIZE - 1);
5714                 src_off_in_page = (start_offset + src_end) &
5715                         (PAGE_CACHE_SIZE - 1);
5716
5717                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5718                 cur = min(cur, dst_off_in_page + 1);
5719                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5720                            dst_off_in_page - cur + 1,
5721                            src_off_in_page - cur + 1, cur);
5722
5723                 dst_end -= cur;
5724                 src_end -= cur;
5725                 len -= cur;
5726         }
5727 }
5728
5729 int try_release_extent_buffer(struct page *page)
5730 {
5731         struct extent_buffer *eb;
5732
5733         /*
5734          * We need to make sure noboody is attaching this page to an eb right
5735          * now.
5736          */
5737         spin_lock(&page->mapping->private_lock);
5738         if (!PagePrivate(page)) {
5739                 spin_unlock(&page->mapping->private_lock);
5740                 return 1;
5741         }
5742
5743         eb = (struct extent_buffer *)page->private;
5744         BUG_ON(!eb);
5745
5746         /*
5747          * This is a little awful but should be ok, we need to make sure that
5748          * the eb doesn't disappear out from under us while we're looking at
5749          * this page.
5750          */
5751         spin_lock(&eb->refs_lock);
5752         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5753                 spin_unlock(&eb->refs_lock);
5754                 spin_unlock(&page->mapping->private_lock);
5755                 return 0;
5756         }
5757         spin_unlock(&page->mapping->private_lock);
5758
5759         /*
5760          * If tree ref isn't set then we know the ref on this eb is a real ref,
5761          * so just return, this page will likely be freed soon anyway.
5762          */
5763         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5764                 spin_unlock(&eb->refs_lock);
5765                 return 0;
5766         }
5767
5768         return release_extent_buffer(eb);
5769 }