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 #include "transaction.h"
  24
  25 static struct kmem_cache *extent_state_cache;
  26 static struct kmem_cache *extent_buffer_cache;
  27 static struct bio_set *btrfs_bioset;
  28
  29 static inline bool extent_state_in_tree(const struct extent_state *state)
  30 {
  31         return !RB_EMPTY_NODE(&state->rb_node);
  32 }
  33
  34 #ifdef CONFIG_BTRFS_DEBUG
  35 static LIST_HEAD(buffers);
  36 static LIST_HEAD(states);
  37
  38 static DEFINE_SPINLOCK(leak_lock);
  39
  40 static inline
  41 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  42 {
  43         unsigned long flags;
  44
  45         spin_lock_irqsave(&leak_lock, flags);
  46         list_add(new, head);
  47         spin_unlock_irqrestore(&leak_lock, flags);
  48 }
  49
  50 static inline
  51 void btrfs_leak_debug_del(struct list_head *entry)
  52 {
  53         unsigned long flags;
  54
  55         spin_lock_irqsave(&leak_lock, flags);
  56         list_del(entry);
  57         spin_unlock_irqrestore(&leak_lock, flags);
  58 }
  59
  60 static inline
  61 void btrfs_leak_debug_check(void)
  62 {
  63         struct extent_state *state;
  64         struct extent_buffer *eb;
  65
  66         while (!list_empty(&states)) {
  67                 state = list_entry(states.next, struct extent_state, leak_list);
  68                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  69                        state->start, state->end, state->state,
  70                        extent_state_in_tree(state),
  71                        atomic_read(&state->refs));
  72                 list_del(&state->leak_list);
  73                 kmem_cache_free(extent_state_cache, state);
  74         }
  75
  76         while (!list_empty(&buffers)) {
  77                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  78                 pr_err("BTRFS: buffer leak start %llu len %lu 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(BTRFS_I(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 REQ_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_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_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         kmem_cache_destroy(extent_state_cache);
 210         kmem_cache_destroy(extent_buffer_cache);
 211         if (btrfs_bioset)
 212                 bioset_free(btrfs_bioset);
 213 }
 214
 215 void extent_io_tree_init(struct extent_io_tree *tree,
 216                          struct address_space *mapping)
 217 {
 218         tree->state = RB_ROOT;
 219         tree->ops = NULL;
 220         tree->dirty_bytes = 0;
 221         spin_lock_init(&tree->lock);
 222         tree->mapping = mapping;
 223 }
 224
 225 static struct extent_state *alloc_extent_state(gfp_t mask)
 226 {
 227         struct extent_state *state;
 228
 229         /*
 230          * The given mask might be not appropriate for the slab allocator,
 231          * drop the unsupported bits
 232          */
 233         mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
 234         state = kmem_cache_alloc(extent_state_cache, mask);
 235         if (!state)
 236                 return state;
 237         state->state = 0;
 238         state->failrec = NULL;
 239         RB_CLEAR_NODE(&state->rb_node);
 240         btrfs_leak_debug_add(&state->leak_list, &states);
 241         atomic_set(&state->refs, 1);
 242         init_waitqueue_head(&state->wq);
 243         trace_alloc_extent_state(state, mask, _RET_IP_);
 244         return state;
 245 }
 246
 247 void free_extent_state(struct extent_state *state)
 248 {
 249         if (!state)
 250                 return;
 251         if (atomic_dec_and_test(&state->refs)) {
 252                 WARN_ON(extent_state_in_tree(state));
 253                 btrfs_leak_debug_del(&state->leak_list);
 254                 trace_free_extent_state(state, _RET_IP_);
 255                 kmem_cache_free(extent_state_cache, state);
 256         }
 257 }
 258
 259 static struct rb_node *tree_insert(struct rb_root *root,
 260                                    struct rb_node *search_start,
 261                                    u64 offset,
 262                                    struct rb_node *node,
 263                                    struct rb_node ***p_in,
 264                                    struct rb_node **parent_in)
 265 {
 266         struct rb_node **p;
 267         struct rb_node *parent = NULL;
 268         struct tree_entry *entry;
 269
 270         if (p_in && parent_in) {
 271                 p = *p_in;
 272                 parent = *parent_in;
 273                 goto do_insert;
 274         }
 275
 276         p = search_start ? &search_start : &root->rb_node;
 277         while (*p) {
 278                 parent = *p;
 279                 entry = rb_entry(parent, struct tree_entry, rb_node);
 280
 281                 if (offset < entry->start)
 282                         p = &(*p)->rb_left;
 283                 else if (offset > entry->end)
 284                         p = &(*p)->rb_right;
 285                 else
 286                         return parent;
 287         }
 288
 289 do_insert:
 290         rb_link_node(node, parent, p);
 291         rb_insert_color(node, root);
 292         return NULL;
 293 }
 294
 295 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 296                                       struct rb_node **prev_ret,
 297                                       struct rb_node **next_ret,
 298                                       struct rb_node ***p_ret,
 299                                       struct rb_node **parent_ret)
 300 {
 301         struct rb_root *root = &tree->state;
 302         struct rb_node **n = &root->rb_node;
 303         struct rb_node *prev = NULL;
 304         struct rb_node *orig_prev = NULL;
 305         struct tree_entry *entry;
 306         struct tree_entry *prev_entry = NULL;
 307
 308         while (*n) {
 309                 prev = *n;
 310                 entry = rb_entry(prev, struct tree_entry, rb_node);
 311                 prev_entry = entry;
 312
 313                 if (offset < entry->start)
 314                         n = &(*n)->rb_left;
 315                 else if (offset > entry->end)
 316                         n = &(*n)->rb_right;
 317                 else
 318                         return *n;
 319         }
 320
 321         if (p_ret)
 322                 *p_ret = n;
 323         if (parent_ret)
 324                 *parent_ret = prev;
 325
 326         if (prev_ret) {
 327                 orig_prev = prev;
 328                 while (prev && offset > prev_entry->end) {
 329                         prev = rb_next(prev);
 330                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 331                 }
 332                 *prev_ret = prev;
 333                 prev = orig_prev;
 334         }
 335
 336         if (next_ret) {
 337                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 338                 while (prev && offset < prev_entry->start) {
 339                         prev = rb_prev(prev);
 340                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 341                 }
 342                 *next_ret = prev;
 343         }
 344         return NULL;
 345 }
 346
 347 static inline struct rb_node *
 348 tree_search_for_insert(struct extent_io_tree *tree,
 349                        u64 offset,
 350                        struct rb_node ***p_ret,
 351                        struct rb_node **parent_ret)
 352 {
 353         struct rb_node *prev = NULL;
 354         struct rb_node *ret;
 355
 356         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 357         if (!ret)
 358                 return prev;
 359         return ret;
 360 }
 361
 362 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 363                                           u64 offset)
 364 {
 365         return tree_search_for_insert(tree, offset, NULL, NULL);
 366 }
 367
 368 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 369                      struct extent_state *other)
 370 {
 371         if (tree->ops && tree->ops->merge_extent_hook)
 372                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 373                                              other);
 374 }
 375
 376 /*
 377  * utility function to look for merge candidates inside a given range.
 378  * Any extents with matching state are merged together into a single
 379  * extent in the tree.  Extents with EXTENT_IO in their state field
 380  * are not merged because the end_io handlers need to be able to do
 381  * operations on them without sleeping (or doing allocations/splits).
 382  *
 383  * This should be called with the tree lock held.
 384  */
 385 static void merge_state(struct extent_io_tree *tree,
 386                         struct extent_state *state)
 387 {
 388         struct extent_state *other;
 389         struct rb_node *other_node;
 390
 391         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 392                 return;
 393
 394         other_node = rb_prev(&state->rb_node);
 395         if (other_node) {
 396                 other = rb_entry(other_node, struct extent_state, rb_node);
 397                 if (other->end == state->start - 1 &&
 398                     other->state == state->state) {
 399                         merge_cb(tree, state, other);
 400                         state->start = other->start;
 401                         rb_erase(&other->rb_node, &tree->state);
 402                         RB_CLEAR_NODE(&other->rb_node);
 403                         free_extent_state(other);
 404                 }
 405         }
 406         other_node = rb_next(&state->rb_node);
 407         if (other_node) {
 408                 other = rb_entry(other_node, struct extent_state, rb_node);
 409                 if (other->start == state->end + 1 &&
 410                     other->state == state->state) {
 411                         merge_cb(tree, state, other);
 412                         state->end = other->end;
 413                         rb_erase(&other->rb_node, &tree->state);
 414                         RB_CLEAR_NODE(&other->rb_node);
 415                         free_extent_state(other);
 416                 }
 417         }
 418 }
 419
 420 static void set_state_cb(struct extent_io_tree *tree,
 421                          struct extent_state *state, unsigned *bits)
 422 {
 423         if (tree->ops && tree->ops->set_bit_hook)
 424                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 425 }
 426
 427 static void clear_state_cb(struct extent_io_tree *tree,
 428                            struct extent_state *state, unsigned *bits)
 429 {
 430         if (tree->ops && tree->ops->clear_bit_hook)
 431                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 432 }
 433
 434 static void set_state_bits(struct extent_io_tree *tree,
 435                            struct extent_state *state, unsigned *bits,
 436                            struct extent_changeset *changeset);
 437
 438 /*
 439  * insert an extent_state struct into the tree.  'bits' are set on the
 440  * struct before it is inserted.
 441  *
 442  * This may return -EEXIST if the extent is already there, in which case the
 443  * state struct is freed.
 444  *
 445  * The tree lock is not taken internally.  This is a utility function and
 446  * probably isn't what you want to call (see set/clear_extent_bit).
 447  */
 448 static int insert_state(struct extent_io_tree *tree,
 449                         struct extent_state *state, u64 start, u64 end,
 450                         struct rb_node ***p,
 451                         struct rb_node **parent,
 452                         unsigned *bits, struct extent_changeset *changeset)
 453 {
 454         struct rb_node *node;
 455
 456         if (end < start)
 457                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 458                        end, start);
 459         state->start = start;
 460         state->end = end;
 461
 462         set_state_bits(tree, state, bits, changeset);
 463
 464         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 465         if (node) {
 466                 struct extent_state *found;
 467                 found = rb_entry(node, struct extent_state, rb_node);
 468                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
 469                        found->start, found->end, start, end);
 470                 return -EEXIST;
 471         }
 472         merge_state(tree, state);
 473         return 0;
 474 }
 475
 476 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 477                      u64 split)
 478 {
 479         if (tree->ops && tree->ops->split_extent_hook)
 480                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 481 }
 482
 483 /*
 484  * split a given extent state struct in two, inserting the preallocated
 485  * struct 'prealloc' as the newly created second half.  'split' indicates an
 486  * offset inside 'orig' where it should be split.
 487  *
 488  * Before calling,
 489  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 490  * are two extent state structs in the tree:
 491  * prealloc: [orig->start, split - 1]
 492  * orig: [ split, orig->end ]
 493  *
 494  * The tree locks are not taken by this function. They need to be held
 495  * by the caller.
 496  */
 497 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 498                        struct extent_state *prealloc, u64 split)
 499 {
 500         struct rb_node *node;
 501
 502         split_cb(tree, orig, split);
 503
 504         prealloc->start = orig->start;
 505         prealloc->end = split - 1;
 506         prealloc->state = orig->state;
 507         orig->start = split;
 508
 509         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 510                            &prealloc->rb_node, NULL, NULL);
 511         if (node) {
 512                 free_extent_state(prealloc);
 513                 return -EEXIST;
 514         }
 515         return 0;
 516 }
 517
 518 static struct extent_state *next_state(struct extent_state *state)
 519 {
 520         struct rb_node *next = rb_next(&state->rb_node);
 521         if (next)
 522                 return rb_entry(next, struct extent_state, rb_node);
 523         else
 524                 return NULL;
 525 }
 526
 527 /*
 528  * utility function to clear some bits in an extent state struct.
 529  * it will optionally wake up any one waiting on this state (wake == 1).
 530  *
 531  * If no bits are set on the state struct after clearing things, the
 532  * struct is freed and removed from the tree
 533  */
 534 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 535                                             struct extent_state *state,
 536                                             unsigned *bits, int wake,
 537                                             struct extent_changeset *changeset)
 538 {
 539         struct extent_state *next;
 540         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 541
 542         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 543                 u64 range = state->end - state->start + 1;
 544                 WARN_ON(range > tree->dirty_bytes);
 545                 tree->dirty_bytes -= range;
 546         }
 547         clear_state_cb(tree, state, bits);
 548         add_extent_changeset(state, bits_to_clear, changeset, 0);
 549         state->state &= ~bits_to_clear;
 550         if (wake)
 551                 wake_up(&state->wq);
 552         if (state->state == 0) {
 553                 next = next_state(state);
 554                 if (extent_state_in_tree(state)) {
 555                         rb_erase(&state->rb_node, &tree->state);
 556                         RB_CLEAR_NODE(&state->rb_node);
 557                         free_extent_state(state);
 558                 } else {
 559                         WARN_ON(1);
 560                 }
 561         } else {
 562                 merge_state(tree, state);
 563                 next = next_state(state);
 564         }
 565         return next;
 566 }
 567
 568 static struct extent_state *
 569 alloc_extent_state_atomic(struct extent_state *prealloc)
 570 {
 571         if (!prealloc)
 572                 prealloc = alloc_extent_state(GFP_ATOMIC);
 573
 574         return prealloc;
 575 }
 576
 577 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 578 {
 579         btrfs_panic(tree_fs_info(tree), err,
 580                     "Locking error: Extent tree was modified by another thread while locked.");
 581 }
 582
 583 /*
 584  * clear some bits on a range in the tree.  This may require splitting
 585  * or inserting elements in the tree, so the gfp mask is used to
 586  * indicate which allocations or sleeping are allowed.
 587  *
 588  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 589  * the given range from the tree regardless of state (ie for truncate).
 590  *
 591  * the range [start, end] is inclusive.
 592  *
 593  * This takes the tree lock, and returns 0 on success and < 0 on error.
 594  */
 595 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 596                               unsigned bits, int wake, int delete,
 597                               struct extent_state **cached_state,
 598                               gfp_t mask, struct extent_changeset *changeset)
 599 {
 600         struct extent_state *state;
 601         struct extent_state *cached;
 602         struct extent_state *prealloc = NULL;
 603         struct rb_node *node;
 604         u64 last_end;
 605         int err;
 606         int clear = 0;
 607
 608         btrfs_debug_check_extent_io_range(tree, start, end);
 609
 610         if (bits & EXTENT_DELALLOC)
 611                 bits |= EXTENT_NORESERVE;
 612
 613         if (delete)
 614                 bits |= ~EXTENT_CTLBITS;
 615         bits |= EXTENT_FIRST_DELALLOC;
 616
 617         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 618                 clear = 1;
 619 again:
 620         if (!prealloc && gfpflags_allow_blocking(mask)) {
 621                 /*
 622                  * Don't care for allocation failure here because we might end
 623                  * up not needing the pre-allocated extent state at all, which
 624                  * is the case if we only have in the tree extent states that
 625                  * cover our input range and don't cover too any other range.
 626                  * If we end up needing a new extent state we allocate it later.
 627                  */
 628                 prealloc = alloc_extent_state(mask);
 629         }
 630
 631         spin_lock(&tree->lock);
 632         if (cached_state) {
 633                 cached = *cached_state;
 634
 635                 if (clear) {
 636                         *cached_state = NULL;
 637                         cached_state = NULL;
 638                 }
 639
 640                 if (cached && extent_state_in_tree(cached) &&
 641                     cached->start <= start && cached->end > start) {
 642                         if (clear)
 643                                 atomic_dec(&cached->refs);
 644                         state = cached;
 645                         goto hit_next;
 646                 }
 647                 if (clear)
 648                         free_extent_state(cached);
 649         }
 650         /*
 651          * this search will find the extents that end after
 652          * our range starts
 653          */
 654         node = tree_search(tree, start);
 655         if (!node)
 656                 goto out;
 657         state = rb_entry(node, struct extent_state, rb_node);
 658 hit_next:
 659         if (state->start > end)
 660                 goto out;
 661         WARN_ON(state->end < start);
 662         last_end = state->end;
 663
 664         /* the state doesn't have the wanted bits, go ahead */
 665         if (!(state->state & bits)) {
 666                 state = next_state(state);
 667                 goto next;
 668         }
 669
 670         /*
 671          *     | ---- desired range ---- |
 672          *  | state | or
 673          *  | ------------- state -------------- |
 674          *
 675          * We need to split the extent we found, and may flip
 676          * bits on second half.
 677          *
 678          * If the extent we found extends past our range, we
 679          * just split and search again.  It'll get split again
 680          * the next time though.
 681          *
 682          * If the extent we found is inside our range, we clear
 683          * the desired bit on it.
 684          */
 685
 686         if (state->start < start) {
 687                 prealloc = alloc_extent_state_atomic(prealloc);
 688                 BUG_ON(!prealloc);
 689                 err = split_state(tree, state, prealloc, start);
 690                 if (err)
 691                         extent_io_tree_panic(tree, err);
 692
 693                 prealloc = NULL;
 694                 if (err)
 695                         goto out;
 696                 if (state->end <= end) {
 697                         state = clear_state_bit(tree, state, &bits, wake,
 698                                                 changeset);
 699                         goto next;
 700                 }
 701                 goto search_again;
 702         }
 703         /*
 704          * | ---- desired range ---- |
 705          *                        | state |
 706          * We need to split the extent, and clear the bit
 707          * on the first half
 708          */
 709         if (state->start <= end && state->end > end) {
 710                 prealloc = alloc_extent_state_atomic(prealloc);
 711                 BUG_ON(!prealloc);
 712                 err = split_state(tree, state, prealloc, end + 1);
 713                 if (err)
 714                         extent_io_tree_panic(tree, err);
 715
 716                 if (wake)
 717                         wake_up(&state->wq);
 718
 719                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 720
 721                 prealloc = NULL;
 722                 goto out;
 723         }
 724
 725         state = clear_state_bit(tree, state, &bits, wake, changeset);
 726 next:
 727         if (last_end == (u64)-1)
 728                 goto out;
 729         start = last_end + 1;
 730         if (start <= end && state && !need_resched())
 731                 goto hit_next;
 732
 733 search_again:
 734         if (start > end)
 735                 goto out;
 736         spin_unlock(&tree->lock);
 737         if (gfpflags_allow_blocking(mask))
 738                 cond_resched();
 739         goto again;
 740
 741 out:
 742         spin_unlock(&tree->lock);
 743         if (prealloc)
 744                 free_extent_state(prealloc);
 745
 746         return 0;
 747
 748 }
 749
 750 static void wait_on_state(struct extent_io_tree *tree,
 751                           struct extent_state *state)
 752                 __releases(tree->lock)
 753                 __acquires(tree->lock)
 754 {
 755         DEFINE_WAIT(wait);
 756         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 757         spin_unlock(&tree->lock);
 758         schedule();
 759         spin_lock(&tree->lock);
 760         finish_wait(&state->wq, &wait);
 761 }
 762
 763 /*
 764  * waits for one or more bits to clear on a range in the state tree.
 765  * The range [start, end] is inclusive.
 766  * The tree lock is taken by this function
 767  */
 768 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 769                             unsigned long bits)
 770 {
 771         struct extent_state *state;
 772         struct rb_node *node;
 773
 774         btrfs_debug_check_extent_io_range(tree, start, end);
 775
 776         spin_lock(&tree->lock);
 777 again:
 778         while (1) {
 779                 /*
 780                  * this search will find all the extents that end after
 781                  * our range starts
 782                  */
 783                 node = tree_search(tree, start);
 784 process_node:
 785                 if (!node)
 786                         break;
 787
 788                 state = rb_entry(node, struct extent_state, rb_node);
 789
 790                 if (state->start > end)
 791                         goto out;
 792
 793                 if (state->state & bits) {
 794                         start = state->start;
 795                         atomic_inc(&state->refs);
 796                         wait_on_state(tree, state);
 797                         free_extent_state(state);
 798                         goto again;
 799                 }
 800                 start = state->end + 1;
 801
 802                 if (start > end)
 803                         break;
 804
 805                 if (!cond_resched_lock(&tree->lock)) {
 806                         node = rb_next(node);
 807                         goto process_node;
 808                 }
 809         }
 810 out:
 811         spin_unlock(&tree->lock);
 812 }
 813
 814 static void set_state_bits(struct extent_io_tree *tree,
 815                            struct extent_state *state,
 816                            unsigned *bits, struct extent_changeset *changeset)
 817 {
 818         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 819
 820         set_state_cb(tree, state, bits);
 821         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 822                 u64 range = state->end - state->start + 1;
 823                 tree->dirty_bytes += range;
 824         }
 825         add_extent_changeset(state, bits_to_set, changeset, 1);
 826         state->state |= bits_to_set;
 827 }
 828
 829 static void cache_state_if_flags(struct extent_state *state,
 830                                  struct extent_state **cached_ptr,
 831                                  unsigned flags)
 832 {
 833         if (cached_ptr && !(*cached_ptr)) {
 834                 if (!flags || (state->state & flags)) {
 835                         *cached_ptr = state;
 836                         atomic_inc(&state->refs);
 837                 }
 838         }
 839 }
 840
 841 static void cache_state(struct extent_state *state,
 842                         struct extent_state **cached_ptr)
 843 {
 844         return cache_state_if_flags(state, cached_ptr,
 845                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 846 }
 847
 848 /*
 849  * set some bits on a range in the tree.  This may require allocations or
 850  * sleeping, so the gfp mask is used to indicate what is allowed.
 851  *
 852  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 853  * part of the range already has the desired bits set.  The start of the
 854  * existing range is returned in failed_start in this case.
 855  *
 856  * [start, end] is inclusive This takes the tree lock.
 857  */
 858
 859 static int __must_check
 860 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 861                  unsigned bits, unsigned exclusive_bits,
 862                  u64 *failed_start, struct extent_state **cached_state,
 863                  gfp_t mask, struct extent_changeset *changeset)
 864 {
 865         struct extent_state *state;
 866         struct extent_state *prealloc = NULL;
 867         struct rb_node *node;
 868         struct rb_node **p;
 869         struct rb_node *parent;
 870         int err = 0;
 871         u64 last_start;
 872         u64 last_end;
 873
 874         btrfs_debug_check_extent_io_range(tree, start, end);
 875
 876         bits |= EXTENT_FIRST_DELALLOC;
 877 again:
 878         if (!prealloc && gfpflags_allow_blocking(mask)) {
 879                 /*
 880                  * Don't care for allocation failure here because we might end
 881                  * up not needing the pre-allocated extent state at all, which
 882                  * is the case if we only have in the tree extent states that
 883                  * cover our input range and don't cover too any other range.
 884                  * If we end up needing a new extent state we allocate it later.
 885                  */
 886                 prealloc = alloc_extent_state(mask);
 887         }
 888
 889         spin_lock(&tree->lock);
 890         if (cached_state && *cached_state) {
 891                 state = *cached_state;
 892                 if (state->start <= start && state->end > start &&
 893                     extent_state_in_tree(state)) {
 894                         node = &state->rb_node;
 895                         goto hit_next;
 896                 }
 897         }
 898         /*
 899          * this search will find all the extents that end after
 900          * our range starts.
 901          */
 902         node = tree_search_for_insert(tree, start, &p, &parent);
 903         if (!node) {
 904                 prealloc = alloc_extent_state_atomic(prealloc);
 905                 BUG_ON(!prealloc);
 906                 err = insert_state(tree, prealloc, start, end,
 907                                    &p, &parent, &bits, changeset);
 908                 if (err)
 909                         extent_io_tree_panic(tree, err);
 910
 911                 cache_state(prealloc, cached_state);
 912                 prealloc = NULL;
 913                 goto out;
 914         }
 915         state = rb_entry(node, struct extent_state, rb_node);
 916 hit_next:
 917         last_start = state->start;
 918         last_end = state->end;
 919
 920         /*
 921          * | ---- desired range ---- |
 922          * | state |
 923          *
 924          * Just lock what we found and keep going
 925          */
 926         if (state->start == start && state->end <= end) {
 927                 if (state->state & exclusive_bits) {
 928                         *failed_start = state->start;
 929                         err = -EEXIST;
 930                         goto out;
 931                 }
 932
 933                 set_state_bits(tree, state, &bits, changeset);
 934                 cache_state(state, cached_state);
 935                 merge_state(tree, state);
 936                 if (last_end == (u64)-1)
 937                         goto out;
 938                 start = last_end + 1;
 939                 state = next_state(state);
 940                 if (start < end && state && state->start == start &&
 941                     !need_resched())
 942                         goto hit_next;
 943                 goto search_again;
 944         }
 945
 946         /*
 947          *     | ---- desired range ---- |
 948          * | state |
 949          *   or
 950          * | ------------- state -------------- |
 951          *
 952          * We need to split the extent we found, and may flip bits on
 953          * second half.
 954          *
 955          * If the extent we found extends past our
 956          * range, we just split and search again.  It'll get split
 957          * again the next time though.
 958          *
 959          * If the extent we found is inside our range, we set the
 960          * desired bit on it.
 961          */
 962         if (state->start < start) {
 963                 if (state->state & exclusive_bits) {
 964                         *failed_start = start;
 965                         err = -EEXIST;
 966                         goto out;
 967                 }
 968
 969                 prealloc = alloc_extent_state_atomic(prealloc);
 970                 BUG_ON(!prealloc);
 971                 err = split_state(tree, state, prealloc, start);
 972                 if (err)
 973                         extent_io_tree_panic(tree, err);
 974
 975                 prealloc = NULL;
 976                 if (err)
 977                         goto out;
 978                 if (state->end <= end) {
 979                         set_state_bits(tree, state, &bits, changeset);
 980                         cache_state(state, cached_state);
 981                         merge_state(tree, state);
 982                         if (last_end == (u64)-1)
 983                                 goto out;
 984                         start = last_end + 1;
 985                         state = next_state(state);
 986                         if (start < end && state && state->start == start &&
 987                             !need_resched())
 988                                 goto hit_next;
 989                 }
 990                 goto search_again;
 991         }
 992         /*
 993          * | ---- desired range ---- |
 994          *     | state | or               | state |
 995          *
 996          * There's a hole, we need to insert something in it and
 997          * ignore the extent we found.
 998          */
 999         if (state->start > start) {
1000                 u64 this_end;
1001                 if (end < last_start)
1002                         this_end = end;
1003                 else
1004                         this_end = last_start - 1;
1005
1006                 prealloc = alloc_extent_state_atomic(prealloc);
1007                 BUG_ON(!prealloc);
1008
1009                 /*
1010                  * Avoid to free 'prealloc' if it can be merged with
1011                  * the later extent.
1012                  */
1013                 err = insert_state(tree, prealloc, start, this_end,
1014                                    NULL, NULL, &bits, changeset);
1015                 if (err)
1016                         extent_io_tree_panic(tree, err);
1017
1018                 cache_state(prealloc, cached_state);
1019                 prealloc = NULL;
1020                 start = this_end + 1;
1021                 goto search_again;
1022         }
1023         /*
1024          * | ---- desired range ---- |
1025          *                        | state |
1026          * We need to split the extent, and set the bit
1027          * on the first half
1028          */
1029         if (state->start <= end && state->end > end) {
1030                 if (state->state & exclusive_bits) {
1031                         *failed_start = start;
1032                         err = -EEXIST;
1033                         goto out;
1034                 }
1035
1036                 prealloc = alloc_extent_state_atomic(prealloc);
1037                 BUG_ON(!prealloc);
1038                 err = split_state(tree, state, prealloc, end + 1);
1039                 if (err)
1040                         extent_io_tree_panic(tree, err);
1041
1042                 set_state_bits(tree, prealloc, &bits, changeset);
1043                 cache_state(prealloc, cached_state);
1044                 merge_state(tree, prealloc);
1045                 prealloc = NULL;
1046                 goto out;
1047         }
1048
1049 search_again:
1050         if (start > end)
1051                 goto out;
1052         spin_unlock(&tree->lock);
1053         if (gfpflags_allow_blocking(mask))
1054                 cond_resched();
1055         goto again;
1056
1057 out:
1058         spin_unlock(&tree->lock);
1059         if (prealloc)
1060                 free_extent_state(prealloc);
1061
1062         return err;
1063
1064 }
1065
1066 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1067                    unsigned bits, u64 * failed_start,
1068                    struct extent_state **cached_state, gfp_t mask)
1069 {
1070         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1071                                 cached_state, mask, NULL);
1072 }
1073
1074
1075 /**
1076  * convert_extent_bit - convert all bits in a given range from one bit to
1077  *                      another
1078  * @tree:       the io tree to search
1079  * @start:      the start offset in bytes
1080  * @end:        the end offset in bytes (inclusive)
1081  * @bits:       the bits to set in this range
1082  * @clear_bits: the bits to clear in this range
1083  * @cached_state:       state that we're going to cache
1084  *
1085  * This will go through and set bits for the given range.  If any states exist
1086  * already in this range they are set with the given bit and cleared of the
1087  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1088  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1089  * boundary bits like LOCK.
1090  *
1091  * All allocations are done with GFP_NOFS.
1092  */
1093 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1094                        unsigned bits, unsigned clear_bits,
1095                        struct extent_state **cached_state)
1096 {
1097         struct extent_state *state;
1098         struct extent_state *prealloc = NULL;
1099         struct rb_node *node;
1100         struct rb_node **p;
1101         struct rb_node *parent;
1102         int err = 0;
1103         u64 last_start;
1104         u64 last_end;
1105         bool first_iteration = true;
1106
1107         btrfs_debug_check_extent_io_range(tree, start, end);
1108
1109 again:
1110         if (!prealloc) {
1111                 /*
1112                  * Best effort, don't worry if extent state allocation fails
1113                  * here for the first iteration. We might have a cached state
1114                  * that matches exactly the target range, in which case no
1115                  * extent state allocations are needed. We'll only know this
1116                  * after locking the tree.
1117                  */
1118                 prealloc = alloc_extent_state(GFP_NOFS);
1119                 if (!prealloc && !first_iteration)
1120                         return -ENOMEM;
1121         }
1122
1123         spin_lock(&tree->lock);
1124         if (cached_state && *cached_state) {
1125                 state = *cached_state;
1126                 if (state->start <= start && state->end > start &&
1127                     extent_state_in_tree(state)) {
1128                         node = &state->rb_node;
1129                         goto hit_next;
1130                 }
1131         }
1132
1133         /*
1134          * this search will find all the extents that end after
1135          * our range starts.
1136          */
1137         node = tree_search_for_insert(tree, start, &p, &parent);
1138         if (!node) {
1139                 prealloc = alloc_extent_state_atomic(prealloc);
1140                 if (!prealloc) {
1141                         err = -ENOMEM;
1142                         goto out;
1143                 }
1144                 err = insert_state(tree, prealloc, start, end,
1145                                    &p, &parent, &bits, NULL);
1146                 if (err)
1147                         extent_io_tree_panic(tree, err);
1148                 cache_state(prealloc, cached_state);
1149                 prealloc = NULL;
1150                 goto out;
1151         }
1152         state = rb_entry(node, struct extent_state, rb_node);
1153 hit_next:
1154         last_start = state->start;
1155         last_end = state->end;
1156
1157         /*
1158          * | ---- desired range ---- |
1159          * | state |
1160          *
1161          * Just lock what we found and keep going
1162          */
1163         if (state->start == start && state->end <= end) {
1164                 set_state_bits(tree, state, &bits, NULL);
1165                 cache_state(state, cached_state);
1166                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1167                 if (last_end == (u64)-1)
1168                         goto out;
1169                 start = last_end + 1;
1170                 if (start < end && state && state->start == start &&
1171                     !need_resched())
1172                         goto hit_next;
1173                 goto search_again;
1174         }
1175
1176         /*
1177          *     | ---- desired range ---- |
1178          * | state |
1179          *   or
1180          * | ------------- state -------------- |
1181          *
1182          * We need to split the extent we found, and may flip bits on
1183          * second half.
1184          *
1185          * If the extent we found extends past our
1186          * range, we just split and search again.  It'll get split
1187          * again the next time though.
1188          *
1189          * If the extent we found is inside our range, we set the
1190          * desired bit on it.
1191          */
1192         if (state->start < start) {
1193                 prealloc = alloc_extent_state_atomic(prealloc);
1194                 if (!prealloc) {
1195                         err = -ENOMEM;
1196                         goto out;
1197                 }
1198                 err = split_state(tree, state, prealloc, start);
1199                 if (err)
1200                         extent_io_tree_panic(tree, err);
1201                 prealloc = NULL;
1202                 if (err)
1203                         goto out;
1204                 if (state->end <= end) {
1205                         set_state_bits(tree, state, &bits, NULL);
1206                         cache_state(state, cached_state);
1207                         state = clear_state_bit(tree, state, &clear_bits, 0,
1208                                                 NULL);
1209                         if (last_end == (u64)-1)
1210                                 goto out;
1211                         start = last_end + 1;
1212                         if (start < end && state && state->start == start &&
1213                             !need_resched())
1214                                 goto hit_next;
1215                 }
1216                 goto search_again;
1217         }
1218         /*
1219          * | ---- desired range ---- |
1220          *     | state | or               | state |
1221          *
1222          * There's a hole, we need to insert something in it and
1223          * ignore the extent we found.
1224          */
1225         if (state->start > start) {
1226                 u64 this_end;
1227                 if (end < last_start)
1228                         this_end = end;
1229                 else
1230                         this_end = last_start - 1;
1231
1232                 prealloc = alloc_extent_state_atomic(prealloc);
1233                 if (!prealloc) {
1234                         err = -ENOMEM;
1235                         goto out;
1236                 }
1237
1238                 /*
1239                  * Avoid to free 'prealloc' if it can be merged with
1240                  * the later extent.
1241                  */
1242                 err = insert_state(tree, prealloc, start, this_end,
1243                                    NULL, NULL, &bits, NULL);
1244                 if (err)
1245                         extent_io_tree_panic(tree, err);
1246                 cache_state(prealloc, cached_state);
1247                 prealloc = NULL;
1248                 start = this_end + 1;
1249                 goto search_again;
1250         }
1251         /*
1252          * | ---- desired range ---- |
1253          *                        | state |
1254          * We need to split the extent, and set the bit
1255          * on the first half
1256          */
1257         if (state->start <= end && state->end > end) {
1258                 prealloc = alloc_extent_state_atomic(prealloc);
1259                 if (!prealloc) {
1260                         err = -ENOMEM;
1261                         goto out;
1262                 }
1263
1264                 err = split_state(tree, state, prealloc, end + 1);
1265                 if (err)
1266                         extent_io_tree_panic(tree, err);
1267
1268                 set_state_bits(tree, prealloc, &bits, NULL);
1269                 cache_state(prealloc, cached_state);
1270                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1271                 prealloc = NULL;
1272                 goto out;
1273         }
1274
1275 search_again:
1276         if (start > end)
1277                 goto out;
1278         spin_unlock(&tree->lock);
1279         cond_resched();
1280         first_iteration = false;
1281         goto again;
1282
1283 out:
1284         spin_unlock(&tree->lock);
1285         if (prealloc)
1286                 free_extent_state(prealloc);
1287
1288         return err;
1289 }
1290
1291 /* wrappers around set/clear extent bit */
1292 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1293                            unsigned bits, struct extent_changeset *changeset)
1294 {
1295         /*
1296          * We don't support EXTENT_LOCKED yet, as current changeset will
1297          * record any bits changed, so for EXTENT_LOCKED case, it will
1298          * either fail with -EEXIST or changeset will record the whole
1299          * range.
1300          */
1301         BUG_ON(bits & EXTENT_LOCKED);
1302
1303         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1304                                 changeset);
1305 }
1306
1307 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1308                      unsigned bits, int wake, int delete,
1309                      struct extent_state **cached, gfp_t mask)
1310 {
1311         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1312                                   cached, mask, NULL);
1313 }
1314
1315 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1316                 unsigned bits, struct extent_changeset *changeset)
1317 {
1318         /*
1319          * Don't support EXTENT_LOCKED case, same reason as
1320          * set_record_extent_bits().
1321          */
1322         BUG_ON(bits & EXTENT_LOCKED);
1323
1324         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1325                                   changeset);
1326 }
1327
1328 /*
1329  * either insert or lock state struct between start and end use mask to tell
1330  * us if waiting is desired.
1331  */
1332 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1333                      struct extent_state **cached_state)
1334 {
1335         int err;
1336         u64 failed_start;
1337
1338         while (1) {
1339                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1340                                        EXTENT_LOCKED, &failed_start,
1341                                        cached_state, GFP_NOFS, NULL);
1342                 if (err == -EEXIST) {
1343                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1344                         start = failed_start;
1345                 } else
1346                         break;
1347                 WARN_ON(start > end);
1348         }
1349         return err;
1350 }
1351
1352 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1353 {
1354         int err;
1355         u64 failed_start;
1356
1357         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1358                                &failed_start, NULL, GFP_NOFS, NULL);
1359         if (err == -EEXIST) {
1360                 if (failed_start > start)
1361                         clear_extent_bit(tree, start, failed_start - 1,
1362                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1363                 return 0;
1364         }
1365         return 1;
1366 }
1367
1368 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1369 {
1370         unsigned long index = start >> PAGE_SHIFT;
1371         unsigned long end_index = end >> PAGE_SHIFT;
1372         struct page *page;
1373
1374         while (index <= end_index) {
1375                 page = find_get_page(inode->i_mapping, index);
1376                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1377                 clear_page_dirty_for_io(page);
1378                 put_page(page);
1379                 index++;
1380         }
1381 }
1382
1383 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1384 {
1385         unsigned long index = start >> PAGE_SHIFT;
1386         unsigned long end_index = end >> PAGE_SHIFT;
1387         struct page *page;
1388
1389         while (index <= end_index) {
1390                 page = find_get_page(inode->i_mapping, index);
1391                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1392                 __set_page_dirty_nobuffers(page);
1393                 account_page_redirty(page);
1394                 put_page(page);
1395                 index++;
1396         }
1397 }
1398
1399 /*
1400  * helper function to set both pages and extents in the tree writeback
1401  */
1402 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1403 {
1404         unsigned long index = start >> PAGE_SHIFT;
1405         unsigned long end_index = end >> PAGE_SHIFT;
1406         struct page *page;
1407
1408         while (index <= end_index) {
1409                 page = find_get_page(tree->mapping, index);
1410                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1411                 set_page_writeback(page);
1412                 put_page(page);
1413                 index++;
1414         }
1415 }
1416
1417 /* find the first state struct with 'bits' set after 'start', and
1418  * return it.  tree->lock must be held.  NULL will returned if
1419  * nothing was found after 'start'
1420  */
1421 static struct extent_state *
1422 find_first_extent_bit_state(struct extent_io_tree *tree,
1423                             u64 start, unsigned bits)
1424 {
1425         struct rb_node *node;
1426         struct extent_state *state;
1427
1428         /*
1429          * this search will find all the extents that end after
1430          * our range starts.
1431          */
1432         node = tree_search(tree, start);
1433         if (!node)
1434                 goto out;
1435
1436         while (1) {
1437                 state = rb_entry(node, struct extent_state, rb_node);
1438                 if (state->end >= start && (state->state & bits))
1439                         return state;
1440
1441                 node = rb_next(node);
1442                 if (!node)
1443                         break;
1444         }
1445 out:
1446         return NULL;
1447 }
1448
1449 /*
1450  * find the first offset in the io tree with 'bits' set. zero is
1451  * returned if we find something, and *start_ret and *end_ret are
1452  * set to reflect the state struct that was found.
1453  *
1454  * If nothing was found, 1 is returned. If found something, return 0.
1455  */
1456 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1457                           u64 *start_ret, u64 *end_ret, unsigned bits,
1458                           struct extent_state **cached_state)
1459 {
1460         struct extent_state *state;
1461         struct rb_node *n;
1462         int ret = 1;
1463
1464         spin_lock(&tree->lock);
1465         if (cached_state && *cached_state) {
1466                 state = *cached_state;
1467                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1468                         n = rb_next(&state->rb_node);
1469                         while (n) {
1470                                 state = rb_entry(n, struct extent_state,
1471                                                  rb_node);
1472                                 if (state->state & bits)
1473                                         goto got_it;
1474                                 n = rb_next(n);
1475                         }
1476                         free_extent_state(*cached_state);
1477                         *cached_state = NULL;
1478                         goto out;
1479                 }
1480                 free_extent_state(*cached_state);
1481                 *cached_state = NULL;
1482         }
1483
1484         state = find_first_extent_bit_state(tree, start, bits);
1485 got_it:
1486         if (state) {
1487                 cache_state_if_flags(state, cached_state, 0);
1488                 *start_ret = state->start;
1489                 *end_ret = state->end;
1490                 ret = 0;
1491         }
1492 out:
1493         spin_unlock(&tree->lock);
1494         return ret;
1495 }
1496
1497 /*
1498  * find a contiguous range of bytes in the file marked as delalloc, not
1499  * more than 'max_bytes'.  start and end are used to return the range,
1500  *
1501  * 1 is returned if we find something, 0 if nothing was in the tree
1502  */
1503 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1504                                         u64 *start, u64 *end, u64 max_bytes,
1505                                         struct extent_state **cached_state)
1506 {
1507         struct rb_node *node;
1508         struct extent_state *state;
1509         u64 cur_start = *start;
1510         u64 found = 0;
1511         u64 total_bytes = 0;
1512
1513         spin_lock(&tree->lock);
1514
1515         /*
1516          * this search will find all the extents that end after
1517          * our range starts.
1518          */
1519         node = tree_search(tree, cur_start);
1520         if (!node) {
1521                 if (!found)
1522                         *end = (u64)-1;
1523                 goto out;
1524         }
1525
1526         while (1) {
1527                 state = rb_entry(node, struct extent_state, rb_node);
1528                 if (found && (state->start != cur_start ||
1529                               (state->state & EXTENT_BOUNDARY))) {
1530                         goto out;
1531                 }
1532                 if (!(state->state & EXTENT_DELALLOC)) {
1533                         if (!found)
1534                                 *end = state->end;
1535                         goto out;
1536                 }
1537                 if (!found) {
1538                         *start = state->start;
1539                         *cached_state = state;
1540                         atomic_inc(&state->refs);
1541                 }
1542                 found++;
1543                 *end = state->end;
1544                 cur_start = state->end + 1;
1545                 node = rb_next(node);
1546                 total_bytes += state->end - state->start + 1;
1547                 if (total_bytes >= max_bytes)
1548                         break;
1549                 if (!node)
1550                         break;
1551         }
1552 out:
1553         spin_unlock(&tree->lock);
1554         return found;
1555 }
1556
1557 static noinline void __unlock_for_delalloc(struct inode *inode,
1558                                            struct page *locked_page,
1559                                            u64 start, u64 end)
1560 {
1561         int ret;
1562         struct page *pages[16];
1563         unsigned long index = start >> PAGE_SHIFT;
1564         unsigned long end_index = end >> PAGE_SHIFT;
1565         unsigned long nr_pages = end_index - index + 1;
1566         int i;
1567
1568         if (index == locked_page->index && end_index == index)
1569                 return;
1570
1571         while (nr_pages > 0) {
1572                 ret = find_get_pages_contig(inode->i_mapping, index,
1573                                      min_t(unsigned long, nr_pages,
1574                                      ARRAY_SIZE(pages)), pages);
1575                 for (i = 0; i < ret; i++) {
1576                         if (pages[i] != locked_page)
1577                                 unlock_page(pages[i]);
1578                         put_page(pages[i]);
1579                 }
1580                 nr_pages -= ret;
1581                 index += ret;
1582                 cond_resched();
1583         }
1584 }
1585
1586 static noinline int lock_delalloc_pages(struct inode *inode,
1587                                         struct page *locked_page,
1588                                         u64 delalloc_start,
1589                                         u64 delalloc_end)
1590 {
1591         unsigned long index = delalloc_start >> PAGE_SHIFT;
1592         unsigned long start_index = index;
1593         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1594         unsigned long pages_locked = 0;
1595         struct page *pages[16];
1596         unsigned long nrpages;
1597         int ret;
1598         int i;
1599
1600         /* the caller is responsible for locking the start index */
1601         if (index == locked_page->index && index == end_index)
1602                 return 0;
1603
1604         /* skip the page at the start index */
1605         nrpages = end_index - index + 1;
1606         while (nrpages > 0) {
1607                 ret = find_get_pages_contig(inode->i_mapping, index,
1608                                      min_t(unsigned long,
1609                                      nrpages, ARRAY_SIZE(pages)), pages);
1610                 if (ret == 0) {
1611                         ret = -EAGAIN;
1612                         goto done;
1613                 }
1614                 /* now we have an array of pages, lock them all */
1615                 for (i = 0; i < ret; i++) {
1616                         /*
1617                          * the caller is taking responsibility for
1618                          * locked_page
1619                          */
1620                         if (pages[i] != locked_page) {
1621                                 lock_page(pages[i]);
1622                                 if (!PageDirty(pages[i]) ||
1623                                     pages[i]->mapping != inode->i_mapping) {
1624                                         ret = -EAGAIN;
1625                                         unlock_page(pages[i]);
1626                                         put_page(pages[i]);
1627                                         goto done;
1628                                 }
1629                         }
1630                         put_page(pages[i]);
1631                         pages_locked++;
1632                 }
1633                 nrpages -= ret;
1634                 index += ret;
1635                 cond_resched();
1636         }
1637         ret = 0;
1638 done:
1639         if (ret && pages_locked) {
1640                 __unlock_for_delalloc(inode, locked_page,
1641                               delalloc_start,
1642                               ((u64)(start_index + pages_locked - 1)) <<
1643                               PAGE_SHIFT);
1644         }
1645         return ret;
1646 }
1647
1648 /*
1649  * find a contiguous range of bytes in the file marked as delalloc, not
1650  * more than 'max_bytes'.  start and end are used to return the range,
1651  *
1652  * 1 is returned if we find something, 0 if nothing was in the tree
1653  */
1654 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1655                                     struct extent_io_tree *tree,
1656                                     struct page *locked_page, u64 *start,
1657                                     u64 *end, u64 max_bytes)
1658 {
1659         u64 delalloc_start;
1660         u64 delalloc_end;
1661         u64 found;
1662         struct extent_state *cached_state = NULL;
1663         int ret;
1664         int loops = 0;
1665
1666 again:
1667         /* step one, find a bunch of delalloc bytes starting at start */
1668         delalloc_start = *start;
1669         delalloc_end = 0;
1670         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1671                                     max_bytes, &cached_state);
1672         if (!found || delalloc_end <= *start) {
1673                 *start = delalloc_start;
1674                 *end = delalloc_end;
1675                 free_extent_state(cached_state);
1676                 return 0;
1677         }
1678
1679         /*
1680          * start comes from the offset of locked_page.  We have to lock
1681          * pages in order, so we can't process delalloc bytes before
1682          * locked_page
1683          */
1684         if (delalloc_start < *start)
1685                 delalloc_start = *start;
1686
1687         /*
1688          * make sure to limit the number of pages we try to lock down
1689          */
1690         if (delalloc_end + 1 - delalloc_start > max_bytes)
1691                 delalloc_end = delalloc_start + max_bytes - 1;
1692
1693         /* step two, lock all the pages after the page that has start */
1694         ret = lock_delalloc_pages(inode, locked_page,
1695                                   delalloc_start, delalloc_end);
1696         if (ret == -EAGAIN) {
1697                 /* some of the pages are gone, lets avoid looping by
1698                  * shortening the size of the delalloc range we're searching
1699                  */
1700                 free_extent_state(cached_state);
1701                 cached_state = NULL;
1702                 if (!loops) {
1703                         max_bytes = PAGE_SIZE;
1704                         loops = 1;
1705                         goto again;
1706                 } else {
1707                         found = 0;
1708                         goto out_failed;
1709                 }
1710         }
1711         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1712
1713         /* step three, lock the state bits for the whole range */
1714         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1715
1716         /* then test to make sure it is all still delalloc */
1717         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1718                              EXTENT_DELALLOC, 1, cached_state);
1719         if (!ret) {
1720                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1721                                      &cached_state, GFP_NOFS);
1722                 __unlock_for_delalloc(inode, locked_page,
1723                               delalloc_start, delalloc_end);
1724                 cond_resched();
1725                 goto again;
1726         }
1727         free_extent_state(cached_state);
1728         *start = delalloc_start;
1729         *end = delalloc_end;
1730 out_failed:
1731         return found;
1732 }
1733
1734 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1735                                  u64 delalloc_end, struct page *locked_page,
1736                                  unsigned clear_bits,
1737                                  unsigned long page_ops)
1738 {
1739         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1740         int ret;
1741         struct page *pages[16];
1742         unsigned long index = start >> PAGE_SHIFT;
1743         unsigned long end_index = end >> PAGE_SHIFT;
1744         unsigned long nr_pages = end_index - index + 1;
1745         int i;
1746
1747         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1748         if (page_ops == 0)
1749                 return;
1750
1751         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1752                 mapping_set_error(inode->i_mapping, -EIO);
1753
1754         while (nr_pages > 0) {
1755                 ret = find_get_pages_contig(inode->i_mapping, index,
1756                                      min_t(unsigned long,
1757                                      nr_pages, ARRAY_SIZE(pages)), pages);
1758                 for (i = 0; i < ret; i++) {
1759
1760                         if (page_ops & PAGE_SET_PRIVATE2)
1761                                 SetPagePrivate2(pages[i]);
1762
1763                         if (pages[i] == locked_page) {
1764                                 put_page(pages[i]);
1765                                 continue;
1766                         }
1767                         if (page_ops & PAGE_CLEAR_DIRTY)
1768                                 clear_page_dirty_for_io(pages[i]);
1769                         if (page_ops & PAGE_SET_WRITEBACK)
1770                                 set_page_writeback(pages[i]);
1771                         if (page_ops & PAGE_SET_ERROR)
1772                                 SetPageError(pages[i]);
1773                         if (page_ops & PAGE_END_WRITEBACK)
1774                                 end_page_writeback(pages[i]);
1775                         if (page_ops & PAGE_UNLOCK)
1776                                 unlock_page(pages[i]);
1777                         put_page(pages[i]);
1778                 }
1779                 nr_pages -= ret;
1780                 index += ret;
1781                 cond_resched();
1782         }
1783 }
1784
1785 /*
1786  * count the number of bytes in the tree that have a given bit(s)
1787  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1788  * cached.  The total number found is returned.
1789  */
1790 u64 count_range_bits(struct extent_io_tree *tree,
1791                      u64 *start, u64 search_end, u64 max_bytes,
1792                      unsigned bits, int contig)
1793 {
1794         struct rb_node *node;
1795         struct extent_state *state;
1796         u64 cur_start = *start;
1797         u64 total_bytes = 0;
1798         u64 last = 0;
1799         int found = 0;
1800
1801         if (WARN_ON(search_end <= cur_start))
1802                 return 0;
1803
1804         spin_lock(&tree->lock);
1805         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1806                 total_bytes = tree->dirty_bytes;
1807                 goto out;
1808         }
1809         /*
1810          * this search will find all the extents that end after
1811          * our range starts.
1812          */
1813         node = tree_search(tree, cur_start);
1814         if (!node)
1815                 goto out;
1816
1817         while (1) {
1818                 state = rb_entry(node, struct extent_state, rb_node);
1819                 if (state->start > search_end)
1820                         break;
1821                 if (contig && found && state->start > last + 1)
1822                         break;
1823                 if (state->end >= cur_start && (state->state & bits) == bits) {
1824                         total_bytes += min(search_end, state->end) + 1 -
1825                                        max(cur_start, state->start);
1826                         if (total_bytes >= max_bytes)
1827                                 break;
1828                         if (!found) {
1829                                 *start = max(cur_start, state->start);
1830                                 found = 1;
1831                         }
1832                         last = state->end;
1833                 } else if (contig && found) {
1834                         break;
1835                 }
1836                 node = rb_next(node);
1837                 if (!node)
1838                         break;
1839         }
1840 out:
1841         spin_unlock(&tree->lock);
1842         return total_bytes;
1843 }
1844
1845 /*
1846  * set the private field for a given byte offset in the tree.  If there isn't
1847  * an extent_state there already, this does nothing.
1848  */
1849 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1850                 struct io_failure_record *failrec)
1851 {
1852         struct rb_node *node;
1853         struct extent_state *state;
1854         int ret = 0;
1855
1856         spin_lock(&tree->lock);
1857         /*
1858          * this search will find all the extents that end after
1859          * our range starts.
1860          */
1861         node = tree_search(tree, start);
1862         if (!node) {
1863                 ret = -ENOENT;
1864                 goto out;
1865         }
1866         state = rb_entry(node, struct extent_state, rb_node);
1867         if (state->start != start) {
1868                 ret = -ENOENT;
1869                 goto out;
1870         }
1871         state->failrec = failrec;
1872 out:
1873         spin_unlock(&tree->lock);
1874         return ret;
1875 }
1876
1877 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1878                 struct io_failure_record **failrec)
1879 {
1880         struct rb_node *node;
1881         struct extent_state *state;
1882         int ret = 0;
1883
1884         spin_lock(&tree->lock);
1885         /*
1886          * this search will find all the extents that end after
1887          * our range starts.
1888          */
1889         node = tree_search(tree, start);
1890         if (!node) {
1891                 ret = -ENOENT;
1892                 goto out;
1893         }
1894         state = rb_entry(node, struct extent_state, rb_node);
1895         if (state->start != start) {
1896                 ret = -ENOENT;
1897                 goto out;
1898         }
1899         *failrec = state->failrec;
1900 out:
1901         spin_unlock(&tree->lock);
1902         return ret;
1903 }
1904
1905 /*
1906  * searches a range in the state tree for a given mask.
1907  * If 'filled' == 1, this returns 1 only if every extent in the tree
1908  * has the bits set.  Otherwise, 1 is returned if any bit in the
1909  * range is found set.
1910  */
1911 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1912                    unsigned bits, int filled, struct extent_state *cached)
1913 {
1914         struct extent_state *state = NULL;
1915         struct rb_node *node;
1916         int bitset = 0;
1917
1918         spin_lock(&tree->lock);
1919         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1920             cached->end > start)
1921                 node = &cached->rb_node;
1922         else
1923                 node = tree_search(tree, start);
1924         while (node && start <= end) {
1925                 state = rb_entry(node, struct extent_state, rb_node);
1926
1927                 if (filled && state->start > start) {
1928                         bitset = 0;
1929                         break;
1930                 }
1931
1932                 if (state->start > end)
1933                         break;
1934
1935                 if (state->state & bits) {
1936                         bitset = 1;
1937                         if (!filled)
1938                                 break;
1939                 } else if (filled) {
1940                         bitset = 0;
1941                         break;
1942                 }
1943
1944                 if (state->end == (u64)-1)
1945                         break;
1946
1947                 start = state->end + 1;
1948                 if (start > end)
1949                         break;
1950                 node = rb_next(node);
1951                 if (!node) {
1952                         if (filled)
1953                                 bitset = 0;
1954                         break;
1955                 }
1956         }
1957         spin_unlock(&tree->lock);
1958         return bitset;
1959 }
1960
1961 /*
1962  * helper function to set a given page up to date if all the
1963  * extents in the tree for that page are up to date
1964  */
1965 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1966 {
1967         u64 start = page_offset(page);
1968         u64 end = start + PAGE_SIZE - 1;
1969         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1970                 SetPageUptodate(page);
1971 }
1972
1973 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1974 {
1975         int ret;
1976         int err = 0;
1977         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1978
1979         set_state_failrec(failure_tree, rec->start, NULL);
1980         ret = clear_extent_bits(failure_tree, rec->start,
1981                                 rec->start + rec->len - 1,
1982                                 EXTENT_LOCKED | EXTENT_DIRTY);
1983         if (ret)
1984                 err = ret;
1985
1986         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1987                                 rec->start + rec->len - 1,
1988                                 EXTENT_DAMAGED);
1989         if (ret && !err)
1990                 err = ret;
1991
1992         kfree(rec);
1993         return err;
1994 }
1995
1996 /*
1997  * this bypasses the standard btrfs submit functions deliberately, as
1998  * the standard behavior is to write all copies in a raid setup. here we only
1999  * want to write the one bad copy. so we do the mapping for ourselves and issue
2000  * submit_bio directly.
2001  * to avoid any synchronization issues, wait for the data after writing, which
2002  * actually prevents the read that triggered the error from finishing.
2003  * currently, there can be no more than two copies of every data bit. thus,
2004  * exactly one rewrite is required.
2005  */
2006 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2007                       struct page *page, unsigned int pg_offset, int mirror_num)
2008 {
2009         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2010         struct bio *bio;
2011         struct btrfs_device *dev;
2012         u64 map_length = 0;
2013         u64 sector;
2014         struct btrfs_bio *bbio = NULL;
2015         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2016         int ret;
2017
2018         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2019         BUG_ON(!mirror_num);
2020
2021         /* we can't repair anything in raid56 yet */
2022         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2023                 return 0;
2024
2025         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2026         if (!bio)
2027                 return -EIO;
2028         bio->bi_iter.bi_size = 0;
2029         map_length = length;
2030
2031         /*
2032          * Avoid races with device replace and make sure our bbio has devices
2033          * associated to its stripes that don't go away while we are doing the
2034          * read repair operation.
2035          */
2036         btrfs_bio_counter_inc_blocked(fs_info);
2037         ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2038                               &map_length, &bbio, mirror_num);
2039         if (ret) {
2040                 btrfs_bio_counter_dec(fs_info);
2041                 bio_put(bio);
2042                 return -EIO;
2043         }
2044         BUG_ON(mirror_num != bbio->mirror_num);
2045         sector = bbio->stripes[mirror_num-1].physical >> 9;
2046         bio->bi_iter.bi_sector = sector;
2047         dev = bbio->stripes[mirror_num-1].dev;
2048         btrfs_put_bbio(bbio);
2049         if (!dev || !dev->bdev || !dev->writeable) {
2050                 btrfs_bio_counter_dec(fs_info);
2051                 bio_put(bio);
2052                 return -EIO;
2053         }
2054         bio->bi_bdev = dev->bdev;
2055         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2056         bio_add_page(bio, page, length, pg_offset);
2057
2058         if (btrfsic_submit_bio_wait(bio)) {
2059                 /* try to remap that extent elsewhere? */
2060                 btrfs_bio_counter_dec(fs_info);
2061                 bio_put(bio);
2062                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2063                 return -EIO;
2064         }
2065
2066         btrfs_info_rl_in_rcu(fs_info,
2067                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2068                                   btrfs_ino(BTRFS_I(inode)), start,
2069                                   rcu_str_deref(dev->name), sector);
2070         btrfs_bio_counter_dec(fs_info);
2071         bio_put(bio);
2072         return 0;
2073 }
2074
2075 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2076                          struct extent_buffer *eb, int mirror_num)
2077 {
2078         u64 start = eb->start;
2079         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2080         int ret = 0;
2081
2082         if (fs_info->sb->s_flags & MS_RDONLY)
2083                 return -EROFS;
2084
2085         for (i = 0; i < num_pages; i++) {
2086                 struct page *p = eb->pages[i];
2087
2088                 ret = repair_io_failure(fs_info->btree_inode, start,
2089                                         PAGE_SIZE, start, p,
2090                                         start - page_offset(p), mirror_num);
2091                 if (ret)
2092                         break;
2093                 start += PAGE_SIZE;
2094         }
2095
2096         return ret;
2097 }
2098
2099 /*
2100  * each time an IO finishes, we do a fast check in the IO failure tree
2101  * to see if we need to process or clean up an io_failure_record
2102  */
2103 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2104                      unsigned int pg_offset)
2105 {
2106         u64 private;
2107         struct io_failure_record *failrec;
2108         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2109         struct extent_state *state;
2110         int num_copies;
2111         int ret;
2112
2113         private = 0;
2114         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2115                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2116         if (!ret)
2117                 return 0;
2118
2119         ret = get_state_failrec(&BTRFS_I(inode)->io_failure_tree, start,
2120                         &failrec);
2121         if (ret)
2122                 return 0;
2123
2124         BUG_ON(!failrec->this_mirror);
2125
2126         if (failrec->in_validation) {
2127                 /* there was no real error, just free the record */
2128                 btrfs_debug(fs_info,
2129                         "clean_io_failure: freeing dummy error at %llu",
2130                         failrec->start);
2131                 goto out;
2132         }
2133         if (fs_info->sb->s_flags & MS_RDONLY)
2134                 goto out;
2135
2136         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2137         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2138                                             failrec->start,
2139                                             EXTENT_LOCKED);
2140         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2141
2142         if (state && state->start <= failrec->start &&
2143             state->end >= failrec->start + failrec->len - 1) {
2144                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2145                                               failrec->len);
2146                 if (num_copies > 1)  {
2147                         repair_io_failure(inode, start, failrec->len,
2148                                           failrec->logical, page,
2149                                           pg_offset, failrec->failed_mirror);
2150                 }
2151         }
2152
2153 out:
2154         free_io_failure(inode, failrec);
2155
2156         return 0;
2157 }
2158
2159 /*
2160  * Can be called when
2161  * - hold extent lock
2162  * - under ordered extent
2163  * - the inode is freeing
2164  */
2165 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2166 {
2167         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2168         struct io_failure_record *failrec;
2169         struct extent_state *state, *next;
2170
2171         if (RB_EMPTY_ROOT(&failure_tree->state))
2172                 return;
2173
2174         spin_lock(&failure_tree->lock);
2175         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2176         while (state) {
2177                 if (state->start > end)
2178                         break;
2179
2180                 ASSERT(state->end <= end);
2181
2182                 next = next_state(state);
2183
2184                 failrec = state->failrec;
2185                 free_extent_state(state);
2186                 kfree(failrec);
2187
2188                 state = next;
2189         }
2190         spin_unlock(&failure_tree->lock);
2191 }
2192
2193 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2194                 struct io_failure_record **failrec_ret)
2195 {
2196         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2197         struct io_failure_record *failrec;
2198         struct extent_map *em;
2199         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2200         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2201         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2202         int ret;
2203         u64 logical;
2204
2205         ret = get_state_failrec(failure_tree, start, &failrec);
2206         if (ret) {
2207                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2208                 if (!failrec)
2209                         return -ENOMEM;
2210
2211                 failrec->start = start;
2212                 failrec->len = end - start + 1;
2213                 failrec->this_mirror = 0;
2214                 failrec->bio_flags = 0;
2215                 failrec->in_validation = 0;
2216
2217                 read_lock(&em_tree->lock);
2218                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2219                 if (!em) {
2220                         read_unlock(&em_tree->lock);
2221                         kfree(failrec);
2222                         return -EIO;
2223                 }
2224
2225                 if (em->start > start || em->start + em->len <= start) {
2226                         free_extent_map(em);
2227                         em = NULL;
2228                 }
2229                 read_unlock(&em_tree->lock);
2230                 if (!em) {
2231                         kfree(failrec);
2232                         return -EIO;
2233                 }
2234
2235                 logical = start - em->start;
2236                 logical = em->block_start + logical;
2237                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2238                         logical = em->block_start;
2239                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2240                         extent_set_compress_type(&failrec->bio_flags,
2241                                                  em->compress_type);
2242                 }
2243
2244                 btrfs_debug(fs_info,
2245                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2246                         logical, start, failrec->len);
2247
2248                 failrec->logical = logical;
2249                 free_extent_map(em);
2250
2251                 /* set the bits in the private failure tree */
2252                 ret = set_extent_bits(failure_tree, start, end,
2253                                         EXTENT_LOCKED | EXTENT_DIRTY);
2254                 if (ret >= 0)
2255                         ret = set_state_failrec(failure_tree, start, failrec);
2256                 /* set the bits in the inode's tree */
2257                 if (ret >= 0)
2258                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2259                 if (ret < 0) {
2260                         kfree(failrec);
2261                         return ret;
2262                 }
2263         } else {
2264                 btrfs_debug(fs_info,
2265                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2266                         failrec->logical, failrec->start, failrec->len,
2267                         failrec->in_validation);
2268                 /*
2269                  * when data can be on disk more than twice, add to failrec here
2270                  * (e.g. with a list for failed_mirror) to make
2271                  * clean_io_failure() clean all those errors at once.
2272                  */
2273         }
2274
2275         *failrec_ret = failrec;
2276
2277         return 0;
2278 }
2279
2280 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2281                            struct io_failure_record *failrec, int failed_mirror)
2282 {
2283         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2284         int num_copies;
2285
2286         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2287         if (num_copies == 1) {
2288                 /*
2289                  * we only have a single copy of the data, so don't bother with
2290                  * all the retry and error correction code that follows. no
2291                  * matter what the error is, it is very likely to persist.
2292                  */
2293                 btrfs_debug(fs_info,
2294                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2295                         num_copies, failrec->this_mirror, failed_mirror);
2296                 return 0;
2297         }
2298
2299         /*
2300          * there are two premises:
2301          *      a) deliver good data to the caller
2302          *      b) correct the bad sectors on disk
2303          */
2304         if (failed_bio->bi_vcnt > 1) {
2305                 /*
2306                  * to fulfill b), we need to know the exact failing sectors, as
2307                  * we don't want to rewrite any more than the failed ones. thus,
2308                  * we need separate read requests for the failed bio
2309                  *
2310                  * if the following BUG_ON triggers, our validation request got
2311                  * merged. we need separate requests for our algorithm to work.
2312                  */
2313                 BUG_ON(failrec->in_validation);
2314                 failrec->in_validation = 1;
2315                 failrec->this_mirror = failed_mirror;
2316         } else {
2317                 /*
2318                  * we're ready to fulfill a) and b) alongside. get a good copy
2319                  * of the failed sector and if we succeed, we have setup
2320                  * everything for repair_io_failure to do the rest for us.
2321                  */
2322                 if (failrec->in_validation) {
2323                         BUG_ON(failrec->this_mirror != failed_mirror);
2324                         failrec->in_validation = 0;
2325                         failrec->this_mirror = 0;
2326                 }
2327                 failrec->failed_mirror = failed_mirror;
2328                 failrec->this_mirror++;
2329                 if (failrec->this_mirror == failed_mirror)
2330                         failrec->this_mirror++;
2331         }
2332
2333         if (failrec->this_mirror > num_copies) {
2334                 btrfs_debug(fs_info,
2335                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2336                         num_copies, failrec->this_mirror, failed_mirror);
2337                 return 0;
2338         }
2339
2340         return 1;
2341 }
2342
2343
2344 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2345                                     struct io_failure_record *failrec,
2346                                     struct page *page, int pg_offset, int icsum,
2347                                     bio_end_io_t *endio_func, void *data)
2348 {
2349         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2350         struct bio *bio;
2351         struct btrfs_io_bio *btrfs_failed_bio;
2352         struct btrfs_io_bio *btrfs_bio;
2353
2354         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2355         if (!bio)
2356                 return NULL;
2357
2358         bio->bi_end_io = endio_func;
2359         bio->bi_iter.bi_sector = failrec->logical >> 9;
2360         bio->bi_bdev = fs_info->fs_devices->latest_bdev;
2361         bio->bi_iter.bi_size = 0;
2362         bio->bi_private = data;
2363
2364         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2365         if (btrfs_failed_bio->csum) {
2366                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2367
2368                 btrfs_bio = btrfs_io_bio(bio);
2369                 btrfs_bio->csum = btrfs_bio->csum_inline;
2370                 icsum *= csum_size;
2371                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2372                        csum_size);
2373         }
2374
2375         bio_add_page(bio, page, failrec->len, pg_offset);
2376
2377         return bio;
2378 }
2379
2380 /*
2381  * this is a generic handler for readpage errors (default
2382  * readpage_io_failed_hook). if other copies exist, read those and write back
2383  * good data to the failed position. does not investigate in remapping the
2384  * failed extent elsewhere, hoping the device will be smart enough to do this as
2385  * needed
2386  */
2387
2388 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2389                               struct page *page, u64 start, u64 end,
2390                               int failed_mirror)
2391 {
2392         struct io_failure_record *failrec;
2393         struct inode *inode = page->mapping->host;
2394         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2395         struct bio *bio;
2396         int read_mode = 0;
2397         int ret;
2398
2399         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2400
2401         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2402         if (ret)
2403                 return ret;
2404
2405         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2406         if (!ret) {
2407                 free_io_failure(inode, failrec);
2408                 return -EIO;
2409         }
2410
2411         if (failed_bio->bi_vcnt > 1)
2412                 read_mode |= REQ_FAILFAST_DEV;
2413
2414         phy_offset >>= inode->i_sb->s_blocksize_bits;
2415         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2416                                       start - page_offset(page),
2417                                       (int)phy_offset, failed_bio->bi_end_io,
2418                                       NULL);
2419         if (!bio) {
2420                 free_io_failure(inode, failrec);
2421                 return -EIO;
2422         }
2423         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2424
2425         btrfs_debug(btrfs_sb(inode->i_sb),
2426                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2427                 read_mode, failrec->this_mirror, failrec->in_validation);
2428
2429         ret = tree->ops->submit_bio_hook(inode, bio, failrec->this_mirror,
2430                                          failrec->bio_flags, 0);
2431         if (ret) {
2432                 free_io_failure(inode, failrec);
2433                 bio_put(bio);
2434         }
2435
2436         return ret;
2437 }
2438
2439 /* lots and lots of room for performance fixes in the end_bio funcs */
2440
2441 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2442 {
2443         int uptodate = (err == 0);
2444         struct extent_io_tree *tree;
2445         int ret = 0;
2446
2447         tree = &BTRFS_I(page->mapping->host)->io_tree;
2448
2449         if (tree->ops && tree->ops->writepage_end_io_hook) {
2450                 ret = tree->ops->writepage_end_io_hook(page, start,
2451                                                end, NULL, uptodate);
2452                 if (ret)
2453                         uptodate = 0;
2454         }
2455
2456         if (!uptodate) {
2457                 ClearPageUptodate(page);
2458                 SetPageError(page);
2459                 ret = ret < 0 ? ret : -EIO;
2460                 mapping_set_error(page->mapping, ret);
2461         }
2462 }
2463
2464 /*
2465  * after a writepage IO is done, we need to:
2466  * clear the uptodate bits on error
2467  * clear the writeback bits in the extent tree for this IO
2468  * end_page_writeback if the page has no more pending IO
2469  *
2470  * Scheduling is not allowed, so the extent state tree is expected
2471  * to have one and only one object corresponding to this IO.
2472  */
2473 static void end_bio_extent_writepage(struct bio *bio)
2474 {
2475         struct bio_vec *bvec;
2476         u64 start;
2477         u64 end;
2478         int i;
2479
2480         bio_for_each_segment_all(bvec, bio, i) {
2481                 struct page *page = bvec->bv_page;
2482                 struct inode *inode = page->mapping->host;
2483                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2484
2485                 /* We always issue full-page reads, but if some block
2486                  * in a page fails to read, blk_update_request() will
2487                  * advance bv_offset and adjust bv_len to compensate.
2488                  * Print a warning for nonzero offsets, and an error
2489                  * if they don't add up to a full page.  */
2490                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2491                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2492                                 btrfs_err(fs_info,
2493                                    "partial page write in btrfs with offset %u and length %u",
2494                                         bvec->bv_offset, bvec->bv_len);
2495                         else
2496                                 btrfs_info(fs_info,
2497                                    "incomplete page write in btrfs with offset %u and length %u",
2498                                         bvec->bv_offset, bvec->bv_len);
2499                 }
2500
2501                 start = page_offset(page);
2502                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2503
2504                 end_extent_writepage(page, bio->bi_error, start, end);
2505                 end_page_writeback(page);
2506         }
2507
2508         bio_put(bio);
2509 }
2510
2511 static void
2512 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2513                               int uptodate)
2514 {
2515         struct extent_state *cached = NULL;
2516         u64 end = start + len - 1;
2517
2518         if (uptodate && tree->track_uptodate)
2519                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2520         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2521 }
2522
2523 /*
2524  * after a readpage IO is done, we need to:
2525  * clear the uptodate bits on error
2526  * set the uptodate bits if things worked
2527  * set the page up to date if all extents in the tree are uptodate
2528  * clear the lock bit in the extent tree
2529  * unlock the page if there are no other extents locked for it
2530  *
2531  * Scheduling is not allowed, so the extent state tree is expected
2532  * to have one and only one object corresponding to this IO.
2533  */
2534 static void end_bio_extent_readpage(struct bio *bio)
2535 {
2536         struct bio_vec *bvec;
2537         int uptodate = !bio->bi_error;
2538         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2539         struct extent_io_tree *tree;
2540         u64 offset = 0;
2541         u64 start;
2542         u64 end;
2543         u64 len;
2544         u64 extent_start = 0;
2545         u64 extent_len = 0;
2546         int mirror;
2547         int ret;
2548         int i;
2549
2550         bio_for_each_segment_all(bvec, bio, i) {
2551                 struct page *page = bvec->bv_page;
2552                 struct inode *inode = page->mapping->host;
2553                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2554
2555                 btrfs_debug(fs_info,
2556                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2557                         (u64)bio->bi_iter.bi_sector, bio->bi_error,
2558                         io_bio->mirror_num);
2559                 tree = &BTRFS_I(inode)->io_tree;
2560
2561                 /* We always issue full-page reads, but if some block
2562                  * in a page fails to read, blk_update_request() will
2563                  * advance bv_offset and adjust bv_len to compensate.
2564                  * Print a warning for nonzero offsets, and an error
2565                  * if they don't add up to a full page.  */
2566                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2567                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2568                                 btrfs_err(fs_info,
2569                                         "partial page read in btrfs with offset %u and length %u",
2570                                         bvec->bv_offset, bvec->bv_len);
2571                         else
2572                                 btrfs_info(fs_info,
2573                                         "incomplete page read in btrfs with offset %u and length %u",
2574                                         bvec->bv_offset, bvec->bv_len);
2575                 }
2576
2577                 start = page_offset(page);
2578                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2579                 len = bvec->bv_len;
2580
2581                 mirror = io_bio->mirror_num;
2582                 if (likely(uptodate && tree->ops &&
2583                            tree->ops->readpage_end_io_hook)) {
2584                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2585                                                               page, start, end,
2586                                                               mirror);
2587                         if (ret)
2588                                 uptodate = 0;
2589                         else
2590                                 clean_io_failure(inode, start, page, 0);
2591                 }
2592
2593                 if (likely(uptodate))
2594                         goto readpage_ok;
2595
2596                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2597                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2598                         if (!ret && !bio->bi_error)
2599                                 uptodate = 1;
2600                 } else {
2601                         /*
2602                          * The generic bio_readpage_error handles errors the
2603                          * following way: If possible, new read requests are
2604                          * created and submitted and will end up in
2605                          * end_bio_extent_readpage as well (if we're lucky, not
2606                          * in the !uptodate case). In that case it returns 0 and
2607                          * we just go on with the next page in our bio. If it
2608                          * can't handle the error it will return -EIO and we
2609                          * remain responsible for that page.
2610                          */
2611                         ret = bio_readpage_error(bio, offset, page, start, end,
2612                                                  mirror);
2613                         if (ret == 0) {
2614                                 uptodate = !bio->bi_error;
2615                                 offset += len;
2616                                 continue;
2617                         }
2618                 }
2619 readpage_ok:
2620                 if (likely(uptodate)) {
2621                         loff_t i_size = i_size_read(inode);
2622                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2623                         unsigned off;
2624
2625                         /* Zero out the end if this page straddles i_size */
2626                         off = i_size & (PAGE_SIZE-1);
2627                         if (page->index == end_index && off)
2628                                 zero_user_segment(page, off, PAGE_SIZE);
2629                         SetPageUptodate(page);
2630                 } else {
2631                         ClearPageUptodate(page);
2632                         SetPageError(page);
2633                 }
2634                 unlock_page(page);
2635                 offset += len;
2636
2637                 if (unlikely(!uptodate)) {
2638                         if (extent_len) {
2639                                 endio_readpage_release_extent(tree,
2640                                                               extent_start,
2641                                                               extent_len, 1);
2642                                 extent_start = 0;
2643                                 extent_len = 0;
2644                         }
2645                         endio_readpage_release_extent(tree, start,
2646                                                       end - start + 1, 0);
2647                 } else if (!extent_len) {
2648                         extent_start = start;
2649                         extent_len = end + 1 - start;
2650                 } else if (extent_start + extent_len == start) {
2651                         extent_len += end + 1 - start;
2652                 } else {
2653                         endio_readpage_release_extent(tree, extent_start,
2654                                                       extent_len, uptodate);
2655                         extent_start = start;
2656                         extent_len = end + 1 - start;
2657                 }
2658         }
2659
2660         if (extent_len)
2661                 endio_readpage_release_extent(tree, extent_start, extent_len,
2662                                               uptodate);
2663         if (io_bio->end_io)
2664                 io_bio->end_io(io_bio, bio->bi_error);
2665         bio_put(bio);
2666 }
2667
2668 /*
2669  * this allocates from the btrfs_bioset.  We're returning a bio right now
2670  * but you can call btrfs_io_bio for the appropriate container_of magic
2671  */
2672 struct bio *
2673 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2674                 gfp_t gfp_flags)
2675 {
2676         struct btrfs_io_bio *btrfs_bio;
2677         struct bio *bio;
2678
2679         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2680
2681         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2682                 while (!bio && (nr_vecs /= 2)) {
2683                         bio = bio_alloc_bioset(gfp_flags,
2684                                                nr_vecs, btrfs_bioset);
2685                 }
2686         }
2687
2688         if (bio) {
2689                 bio->bi_bdev = bdev;
2690                 bio->bi_iter.bi_sector = first_sector;
2691                 btrfs_bio = btrfs_io_bio(bio);
2692                 btrfs_bio->csum = NULL;
2693                 btrfs_bio->csum_allocated = NULL;
2694                 btrfs_bio->end_io = NULL;
2695         }
2696         return bio;
2697 }
2698
2699 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2700 {
2701         struct btrfs_io_bio *btrfs_bio;
2702         struct bio *new;
2703
2704         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2705         if (new) {
2706                 btrfs_bio = btrfs_io_bio(new);
2707                 btrfs_bio->csum = NULL;
2708                 btrfs_bio->csum_allocated = NULL;
2709                 btrfs_bio->end_io = NULL;
2710         }
2711         return new;
2712 }
2713
2714 /* this also allocates from the btrfs_bioset */
2715 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2716 {
2717         struct btrfs_io_bio *btrfs_bio;
2718         struct bio *bio;
2719
2720         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2721         if (bio) {
2722                 btrfs_bio = btrfs_io_bio(bio);
2723                 btrfs_bio->csum = NULL;
2724                 btrfs_bio->csum_allocated = NULL;
2725                 btrfs_bio->end_io = NULL;
2726         }
2727         return bio;
2728 }
2729
2730
2731 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2732                                        unsigned long bio_flags)
2733 {
2734         int ret = 0;
2735         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2736         struct page *page = bvec->bv_page;
2737         struct extent_io_tree *tree = bio->bi_private;
2738         u64 start;
2739
2740         start = page_offset(page) + bvec->bv_offset;
2741
2742         bio->bi_private = NULL;
2743         bio_get(bio);
2744
2745         if (tree->ops && tree->ops->submit_bio_hook)
2746                 ret = tree->ops->submit_bio_hook(page->mapping->host, bio,
2747                                            mirror_num, bio_flags, start);
2748         else
2749                 btrfsic_submit_bio(bio);
2750
2751         bio_put(bio);
2752         return ret;
2753 }
2754
2755 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2756                      unsigned long offset, size_t size, struct bio *bio,
2757                      unsigned long bio_flags)
2758 {
2759         int ret = 0;
2760         if (tree->ops && tree->ops->merge_bio_hook)
2761                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2762                                                 bio_flags);
2763         return ret;
2764
2765 }
2766
2767 static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree,
2768                               struct writeback_control *wbc,
2769                               struct page *page, sector_t sector,
2770                               size_t size, unsigned long offset,
2771                               struct block_device *bdev,
2772                               struct bio **bio_ret,
2773                               unsigned long max_pages,
2774                               bio_end_io_t end_io_func,
2775                               int mirror_num,
2776                               unsigned long prev_bio_flags,
2777                               unsigned long bio_flags,
2778                               bool force_bio_submit)
2779 {
2780         int ret = 0;
2781         struct bio *bio;
2782         int contig = 0;
2783         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2784         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2785
2786         if (bio_ret && *bio_ret) {
2787                 bio = *bio_ret;
2788                 if (old_compressed)
2789                         contig = bio->bi_iter.bi_sector == sector;
2790                 else
2791                         contig = bio_end_sector(bio) == sector;
2792
2793                 if (prev_bio_flags != bio_flags || !contig ||
2794                     force_bio_submit ||
2795                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2796                     bio_add_page(bio, page, page_size, offset) < page_size) {
2797                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2798                         if (ret < 0) {
2799                                 *bio_ret = NULL;
2800                                 return ret;
2801                         }
2802                         bio = NULL;
2803                 } else {
2804                         if (wbc)
2805                                 wbc_account_io(wbc, page, page_size);
2806                         return 0;
2807                 }
2808         }
2809
2810         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2811                         GFP_NOFS | __GFP_HIGH);
2812         if (!bio)
2813                 return -ENOMEM;
2814
2815         bio_add_page(bio, page, page_size, offset);
2816         bio->bi_end_io = end_io_func;
2817         bio->bi_private = tree;
2818         bio_set_op_attrs(bio, op, op_flags);
2819         if (wbc) {
2820                 wbc_init_bio(wbc, bio);
2821                 wbc_account_io(wbc, page, page_size);
2822         }
2823
2824         if (bio_ret)
2825                 *bio_ret = bio;
2826         else
2827                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2828
2829         return ret;
2830 }
2831
2832 static void attach_extent_buffer_page(struct extent_buffer *eb,
2833                                       struct page *page)
2834 {
2835         if (!PagePrivate(page)) {
2836                 SetPagePrivate(page);
2837                 get_page(page);
2838                 set_page_private(page, (unsigned long)eb);
2839         } else {
2840                 WARN_ON(page->private != (unsigned long)eb);
2841         }
2842 }
2843
2844 void set_page_extent_mapped(struct page *page)
2845 {
2846         if (!PagePrivate(page)) {
2847                 SetPagePrivate(page);
2848                 get_page(page);
2849                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2850         }
2851 }
2852
2853 static struct extent_map *
2854 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2855                  u64 start, u64 len, get_extent_t *get_extent,
2856                  struct extent_map **em_cached)
2857 {
2858         struct extent_map *em;
2859
2860         if (em_cached && *em_cached) {
2861                 em = *em_cached;
2862                 if (extent_map_in_tree(em) && start >= em->start &&
2863                     start < extent_map_end(em)) {
2864                         atomic_inc(&em->refs);
2865                         return em;
2866                 }
2867
2868                 free_extent_map(em);
2869                 *em_cached = NULL;
2870         }
2871
2872         em = get_extent(inode, page, pg_offset, start, len, 0);
2873         if (em_cached && !IS_ERR_OR_NULL(em)) {
2874                 BUG_ON(*em_cached);
2875                 atomic_inc(&em->refs);
2876                 *em_cached = em;
2877         }
2878         return em;
2879 }
2880 /*
2881  * basic readpage implementation.  Locked extent state structs are inserted
2882  * into the tree that are removed when the IO is done (by the end_io
2883  * handlers)
2884  * XXX JDM: This needs looking at to ensure proper page locking
2885  * return 0 on success, otherwise return error
2886  */
2887 static int __do_readpage(struct extent_io_tree *tree,
2888                          struct page *page,
2889                          get_extent_t *get_extent,
2890                          struct extent_map **em_cached,
2891                          struct bio **bio, int mirror_num,
2892                          unsigned long *bio_flags, int read_flags,
2893                          u64 *prev_em_start)
2894 {
2895         struct inode *inode = page->mapping->host;
2896         u64 start = page_offset(page);
2897         u64 page_end = start + PAGE_SIZE - 1;
2898         u64 end;
2899         u64 cur = start;
2900         u64 extent_offset;
2901         u64 last_byte = i_size_read(inode);
2902         u64 block_start;
2903         u64 cur_end;
2904         sector_t sector;
2905         struct extent_map *em;
2906         struct block_device *bdev;
2907         int ret = 0;
2908         int nr = 0;
2909         size_t pg_offset = 0;
2910         size_t iosize;
2911         size_t disk_io_size;
2912         size_t blocksize = inode->i_sb->s_blocksize;
2913         unsigned long this_bio_flag = 0;
2914
2915         set_page_extent_mapped(page);
2916
2917         end = page_end;
2918         if (!PageUptodate(page)) {
2919                 if (cleancache_get_page(page) == 0) {
2920                         BUG_ON(blocksize != PAGE_SIZE);
2921                         unlock_extent(tree, start, end);
2922                         goto out;
2923                 }
2924         }
2925
2926         if (page->index == last_byte >> PAGE_SHIFT) {
2927                 char *userpage;
2928                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2929
2930                 if (zero_offset) {
2931                         iosize = PAGE_SIZE - zero_offset;
2932                         userpage = kmap_atomic(page);
2933                         memset(userpage + zero_offset, 0, iosize);
2934                         flush_dcache_page(page);
2935                         kunmap_atomic(userpage);
2936                 }
2937         }
2938         while (cur <= end) {
2939                 unsigned long pnr = (last_byte >> PAGE_SHIFT) + 1;
2940                 bool force_bio_submit = false;
2941
2942                 if (cur >= last_byte) {
2943                         char *userpage;
2944                         struct extent_state *cached = NULL;
2945
2946                         iosize = PAGE_SIZE - pg_offset;
2947                         userpage = kmap_atomic(page);
2948                         memset(userpage + pg_offset, 0, iosize);
2949                         flush_dcache_page(page);
2950                         kunmap_atomic(userpage);
2951                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2952                                             &cached, GFP_NOFS);
2953                         unlock_extent_cached(tree, cur,
2954                                              cur + iosize - 1,
2955                                              &cached, GFP_NOFS);
2956                         break;
2957                 }
2958                 em = __get_extent_map(inode, page, pg_offset, cur,
2959                                       end - cur + 1, get_extent, em_cached);
2960                 if (IS_ERR_OR_NULL(em)) {
2961                         SetPageError(page);
2962                         unlock_extent(tree, cur, end);
2963                         break;
2964                 }
2965                 extent_offset = cur - em->start;
2966                 BUG_ON(extent_map_end(em) <= cur);
2967                 BUG_ON(end < cur);
2968
2969                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2970                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2971                         extent_set_compress_type(&this_bio_flag,
2972                                                  em->compress_type);
2973                 }
2974
2975                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2976                 cur_end = min(extent_map_end(em) - 1, end);
2977                 iosize = ALIGN(iosize, blocksize);
2978                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2979                         disk_io_size = em->block_len;
2980                         sector = em->block_start >> 9;
2981                 } else {
2982                         sector = (em->block_start + extent_offset) >> 9;
2983                         disk_io_size = iosize;
2984                 }
2985                 bdev = em->bdev;
2986                 block_start = em->block_start;
2987                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2988                         block_start = EXTENT_MAP_HOLE;
2989
2990                 /*
2991                  * If we have a file range that points to a compressed extent
2992                  * and it's followed by a consecutive file range that points to
2993                  * to the same compressed extent (possibly with a different
2994                  * offset and/or length, so it either points to the whole extent
2995                  * or only part of it), we must make sure we do not submit a
2996                  * single bio to populate the pages for the 2 ranges because
2997                  * this makes the compressed extent read zero out the pages
2998                  * belonging to the 2nd range. Imagine the following scenario:
2999                  *
3000                  *  File layout
3001                  *  [0 - 8K]                     [8K - 24K]
3002                  *    |                               |
3003                  *    |                               |
3004                  * points to extent X,         points to extent X,
3005                  * offset 4K, length of 8K     offset 0, length 16K
3006                  *
3007                  * [extent X, compressed length = 4K uncompressed length = 16K]
3008                  *
3009                  * If the bio to read the compressed extent covers both ranges,
3010                  * it will decompress extent X into the pages belonging to the
3011                  * first range and then it will stop, zeroing out the remaining
3012                  * pages that belong to the other range that points to extent X.
3013                  * So here we make sure we submit 2 bios, one for the first
3014                  * range and another one for the third range. Both will target
3015                  * the same physical extent from disk, but we can't currently
3016                  * make the compressed bio endio callback populate the pages
3017                  * for both ranges because each compressed bio is tightly
3018                  * coupled with a single extent map, and each range can have
3019                  * an extent map with a different offset value relative to the
3020                  * uncompressed data of our extent and different lengths. This
3021                  * is a corner case so we prioritize correctness over
3022                  * non-optimal behavior (submitting 2 bios for the same extent).
3023                  */
3024                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3025                     prev_em_start && *prev_em_start != (u64)-1 &&
3026                     *prev_em_start != em->orig_start)
3027                         force_bio_submit = true;
3028
3029                 if (prev_em_start)
3030                         *prev_em_start = em->orig_start;
3031
3032                 free_extent_map(em);
3033                 em = NULL;
3034
3035                 /* we've found a hole, just zero and go on */
3036                 if (block_start == EXTENT_MAP_HOLE) {
3037                         char *userpage;
3038                         struct extent_state *cached = NULL;
3039
3040                         userpage = kmap_atomic(page);
3041                         memset(userpage + pg_offset, 0, iosize);
3042                         flush_dcache_page(page);
3043                         kunmap_atomic(userpage);
3044
3045                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3046                                             &cached, GFP_NOFS);
3047                         unlock_extent_cached(tree, cur,
3048                                              cur + iosize - 1,
3049                                              &cached, GFP_NOFS);
3050                         cur = cur + iosize;
3051                         pg_offset += iosize;
3052                         continue;
3053                 }
3054                 /* the get_extent function already copied into the page */
3055                 if (test_range_bit(tree, cur, cur_end,
3056                                    EXTENT_UPTODATE, 1, NULL)) {
3057                         check_page_uptodate(tree, page);
3058                         unlock_extent(tree, cur, cur + iosize - 1);
3059                         cur = cur + iosize;
3060                         pg_offset += iosize;
3061                         continue;
3062                 }
3063                 /* we have an inline extent but it didn't get marked up
3064                  * to date.  Error out
3065                  */
3066                 if (block_start == EXTENT_MAP_INLINE) {
3067                         SetPageError(page);
3068                         unlock_extent(tree, cur, cur + iosize - 1);
3069                         cur = cur + iosize;
3070                         pg_offset += iosize;
3071                         continue;
3072                 }
3073
3074                 pnr -= page->index;
3075                 ret = submit_extent_page(REQ_OP_READ, read_flags, tree, NULL,
3076                                          page, sector, disk_io_size, pg_offset,
3077                                          bdev, bio, pnr,
3078                                          end_bio_extent_readpage, mirror_num,
3079                                          *bio_flags,
3080                                          this_bio_flag,
3081                                          force_bio_submit);
3082                 if (!ret) {
3083                         nr++;
3084                         *bio_flags = this_bio_flag;
3085                 } else {
3086                         SetPageError(page);
3087                         unlock_extent(tree, cur, cur + iosize - 1);
3088                         goto out;
3089                 }
3090                 cur = cur + iosize;
3091                 pg_offset += iosize;
3092         }
3093 out:
3094         if (!nr) {
3095                 if (!PageError(page))
3096                         SetPageUptodate(page);
3097                 unlock_page(page);
3098         }
3099         return ret;
3100 }
3101
3102 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3103                                              struct page *pages[], int nr_pages,
3104                                              u64 start, u64 end,
3105                                              get_extent_t *get_extent,
3106                                              struct extent_map **em_cached,
3107                                              struct bio **bio, int mirror_num,
3108                                              unsigned long *bio_flags,
3109                                              u64 *prev_em_start)
3110 {
3111         struct inode *inode;
3112         struct btrfs_ordered_extent *ordered;
3113         int index;
3114
3115         inode = pages[0]->mapping->host;
3116         while (1) {
3117                 lock_extent(tree, start, end);
3118                 ordered = btrfs_lookup_ordered_range(inode, start,
3119                                                      end - start + 1);
3120                 if (!ordered)
3121                         break;
3122                 unlock_extent(tree, start, end);
3123                 btrfs_start_ordered_extent(inode, ordered, 1);
3124                 btrfs_put_ordered_extent(ordered);
3125         }
3126
3127         for (index = 0; index < nr_pages; index++) {
3128                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3129                               mirror_num, bio_flags, 0, prev_em_start);
3130                 put_page(pages[index]);
3131         }
3132 }
3133
3134 static void __extent_readpages(struct extent_io_tree *tree,
3135                                struct page *pages[],
3136                                int nr_pages, get_extent_t *get_extent,
3137                                struct extent_map **em_cached,
3138                                struct bio **bio, int mirror_num,
3139                                unsigned long *bio_flags,
3140                                u64 *prev_em_start)
3141 {
3142         u64 start = 0;
3143         u64 end = 0;
3144         u64 page_start;
3145         int index;
3146         int first_index = 0;
3147
3148         for (index = 0; index < nr_pages; index++) {
3149                 page_start = page_offset(pages[index]);
3150                 if (!end) {
3151                         start = page_start;
3152                         end = start + PAGE_SIZE - 1;
3153                         first_index = index;
3154                 } else if (end + 1 == page_start) {
3155                         end += PAGE_SIZE;
3156                 } else {
3157                         __do_contiguous_readpages(tree, &pages[first_index],
3158                                                   index - first_index, start,
3159                                                   end, get_extent, em_cached,
3160                                                   bio, mirror_num, bio_flags,
3161                                                   prev_em_start);
3162                         start = page_start;
3163                         end = start + PAGE_SIZE - 1;
3164                         first_index = index;
3165                 }
3166         }
3167
3168         if (end)
3169                 __do_contiguous_readpages(tree, &pages[first_index],
3170                                           index - first_index, start,
3171                                           end, get_extent, em_cached, bio,
3172                                           mirror_num, bio_flags,
3173                                           prev_em_start);
3174 }
3175
3176 static int __extent_read_full_page(struct extent_io_tree *tree,
3177                                    struct page *page,
3178                                    get_extent_t *get_extent,
3179                                    struct bio **bio, int mirror_num,
3180                                    unsigned long *bio_flags, int read_flags)
3181 {
3182         struct inode *inode = page->mapping->host;
3183         struct btrfs_ordered_extent *ordered;
3184         u64 start = page_offset(page);
3185         u64 end = start + PAGE_SIZE - 1;
3186         int ret;
3187
3188         while (1) {
3189                 lock_extent(tree, start, end);
3190                 ordered = btrfs_lookup_ordered_range(inode, start,
3191                                                 PAGE_SIZE);
3192                 if (!ordered)
3193                         break;
3194                 unlock_extent(tree, start, end);
3195                 btrfs_start_ordered_extent(inode, ordered, 1);
3196                 btrfs_put_ordered_extent(ordered);
3197         }
3198
3199         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3200                             bio_flags, read_flags, NULL);
3201         return ret;
3202 }
3203
3204 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3205                             get_extent_t *get_extent, int mirror_num)
3206 {
3207         struct bio *bio = NULL;
3208         unsigned long bio_flags = 0;
3209         int ret;
3210
3211         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3212                                       &bio_flags, 0);
3213         if (bio)
3214                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3215         return ret;
3216 }
3217
3218 static void update_nr_written(struct page *page, struct writeback_control *wbc,
3219                               unsigned long nr_written)
3220 {
3221         wbc->nr_to_write -= nr_written;
3222 }
3223
3224 /*
3225  * helper for __extent_writepage, doing all of the delayed allocation setup.
3226  *
3227  * This returns 1 if our fill_delalloc function did all the work required
3228  * to write the page (copy into inline extent).  In this case the IO has
3229  * been started and the page is already unlocked.
3230  *
3231  * This returns 0 if all went well (page still locked)
3232  * This returns < 0 if there were errors (page still locked)
3233  */
3234 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3235                               struct page *page, struct writeback_control *wbc,
3236                               struct extent_page_data *epd,
3237                               u64 delalloc_start,
3238                               unsigned long *nr_written)
3239 {
3240         struct extent_io_tree *tree = epd->tree;
3241         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3242         u64 nr_delalloc;
3243         u64 delalloc_to_write = 0;
3244         u64 delalloc_end = 0;
3245         int ret;
3246         int page_started = 0;
3247
3248         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3249                 return 0;
3250
3251         while (delalloc_end < page_end) {
3252                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3253                                                page,
3254                                                &delalloc_start,
3255                                                &delalloc_end,
3256                                                BTRFS_MAX_EXTENT_SIZE);
3257                 if (nr_delalloc == 0) {
3258                         delalloc_start = delalloc_end + 1;
3259                         continue;
3260                 }
3261                 ret = tree->ops->fill_delalloc(inode, page,
3262                                                delalloc_start,
3263                                                delalloc_end,
3264                                                &page_started,
3265                                                nr_written);
3266                 /* File system has been set read-only */
3267                 if (ret) {
3268                         SetPageError(page);
3269                         /* fill_delalloc should be return < 0 for error
3270                          * but just in case, we use > 0 here meaning the
3271                          * IO is started, so we don't want to return > 0
3272                          * unless things are going well.
3273                          */
3274                         ret = ret < 0 ? ret : -EIO;
3275                         goto done;
3276                 }
3277                 /*
3278                  * delalloc_end is already one less than the total length, so
3279                  * we don't subtract one from PAGE_SIZE
3280                  */
3281                 delalloc_to_write += (delalloc_end - delalloc_start +
3282                                       PAGE_SIZE) >> PAGE_SHIFT;
3283                 delalloc_start = delalloc_end + 1;
3284         }
3285         if (wbc->nr_to_write < delalloc_to_write) {
3286                 int thresh = 8192;
3287
3288                 if (delalloc_to_write < thresh * 2)
3289                         thresh = delalloc_to_write;
3290                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3291                                          thresh);
3292         }
3293
3294         /* did the fill delalloc function already unlock and start
3295          * the IO?
3296          */
3297         if (page_started) {
3298                 /*
3299                  * we've unlocked the page, so we can't update
3300                  * the mapping's writeback index, just update
3301                  * nr_to_write.
3302                  */
3303                 wbc->nr_to_write -= *nr_written;
3304                 return 1;
3305         }
3306
3307         ret = 0;
3308
3309 done:
3310         return ret;
3311 }
3312
3313 /*
3314  * helper for __extent_writepage.  This calls the writepage start hooks,
3315  * and does the loop to map the page into extents and bios.
3316  *
3317  * We return 1 if the IO is started and the page is unlocked,
3318  * 0 if all went well (page still locked)
3319  * < 0 if there were errors (page still locked)
3320  */
3321 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3322                                  struct page *page,
3323                                  struct writeback_control *wbc,
3324                                  struct extent_page_data *epd,
3325                                  loff_t i_size,
3326                                  unsigned long nr_written,
3327                                  int write_flags, int *nr_ret)
3328 {
3329         struct extent_io_tree *tree = epd->tree;
3330         u64 start = page_offset(page);
3331         u64 page_end = start + PAGE_SIZE - 1;
3332         u64 end;
3333         u64 cur = start;
3334         u64 extent_offset;
3335         u64 block_start;
3336         u64 iosize;
3337         sector_t sector;
3338         struct extent_state *cached_state = NULL;
3339         struct extent_map *em;
3340         struct block_device *bdev;
3341         size_t pg_offset = 0;
3342         size_t blocksize;
3343         int ret = 0;
3344         int nr = 0;
3345         bool compressed;
3346
3347         if (tree->ops && tree->ops->writepage_start_hook) {
3348                 ret = tree->ops->writepage_start_hook(page, start,
3349                                                       page_end);
3350                 if (ret) {
3351                         /* Fixup worker will requeue */
3352                         if (ret == -EBUSY)
3353                                 wbc->pages_skipped++;
3354                         else
3355                                 redirty_page_for_writepage(wbc, page);
3356
3357                         update_nr_written(page, wbc, nr_written);
3358                         unlock_page(page);
3359                         ret = 1;
3360                         goto done_unlocked;
3361                 }
3362         }
3363
3364         /*
3365          * we don't want to touch the inode after unlocking the page,
3366          * so we update the mapping writeback index now
3367          */
3368         update_nr_written(page, wbc, nr_written + 1);
3369
3370         end = page_end;
3371         if (i_size <= start) {
3372                 if (tree->ops && tree->ops->writepage_end_io_hook)
3373                         tree->ops->writepage_end_io_hook(page, start,
3374                                                          page_end, NULL, 1);
3375                 goto done;
3376         }
3377
3378         blocksize = inode->i_sb->s_blocksize;
3379
3380         while (cur <= end) {
3381                 u64 em_end;
3382                 unsigned long max_nr;
3383
3384                 if (cur >= i_size) {
3385                         if (tree->ops && tree->ops->writepage_end_io_hook)
3386                                 tree->ops->writepage_end_io_hook(page, cur,
3387                                                          page_end, NULL, 1);
3388                         break;
3389                 }
3390                 em = epd->get_extent(inode, page, pg_offset, cur,
3391                                      end - cur + 1, 1);
3392                 if (IS_ERR_OR_NULL(em)) {
3393                         SetPageError(page);
3394                         ret = PTR_ERR_OR_ZERO(em);
3395                         break;
3396                 }
3397
3398                 extent_offset = cur - em->start;
3399                 em_end = extent_map_end(em);
3400                 BUG_ON(em_end <= cur);
3401                 BUG_ON(end < cur);
3402                 iosize = min(em_end - cur, end - cur + 1);
3403                 iosize = ALIGN(iosize, blocksize);
3404                 sector = (em->block_start + extent_offset) >> 9;
3405                 bdev = em->bdev;
3406                 block_start = em->block_start;
3407                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3408                 free_extent_map(em);
3409                 em = NULL;
3410
3411                 /*
3412                  * compressed and inline extents are written through other
3413                  * paths in the FS
3414                  */
3415                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3416                     block_start == EXTENT_MAP_INLINE) {
3417                         /*
3418                          * end_io notification does not happen here for
3419                          * compressed extents
3420                          */
3421                         if (!compressed && tree->ops &&
3422                             tree->ops->writepage_end_io_hook)
3423                                 tree->ops->writepage_end_io_hook(page, cur,
3424                                                          cur + iosize - 1,
3425                                                          NULL, 1);
3426                         else if (compressed) {
3427                                 /* we don't want to end_page_writeback on
3428                                  * a compressed extent.  this happens
3429                                  * elsewhere
3430                                  */
3431                                 nr++;
3432                         }
3433
3434                         cur += iosize;
3435                         pg_offset += iosize;
3436                         continue;
3437                 }
3438
3439                 max_nr = (i_size >> PAGE_SHIFT) + 1;
3440
3441                 set_range_writeback(tree, cur, cur + iosize - 1);
3442                 if (!PageWriteback(page)) {
3443                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3444                                    "page %lu not writeback, cur %llu end %llu",
3445                                page->index, cur, end);
3446                 }
3447
3448                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3449                                          page, sector, iosize, pg_offset,
3450                                          bdev, &epd->bio, max_nr,
3451                                          end_bio_extent_writepage,
3452                                          0, 0, 0, false);
3453                 if (ret)
3454                         SetPageError(page);
3455
3456                 cur = cur + iosize;
3457                 pg_offset += iosize;
3458                 nr++;
3459         }
3460 done:
3461         *nr_ret = nr;
3462
3463 done_unlocked:
3464
3465         /* drop our reference on any cached states */
3466         free_extent_state(cached_state);
3467         return ret;
3468 }
3469
3470 /*
3471  * the writepage semantics are similar to regular writepage.  extent
3472  * records are inserted to lock ranges in the tree, and as dirty areas
3473  * are found, they are marked writeback.  Then the lock bits are removed
3474  * and the end_io handler clears the writeback ranges
3475  */
3476 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3477                               void *data)
3478 {
3479         struct inode *inode = page->mapping->host;
3480         struct extent_page_data *epd = data;
3481         u64 start = page_offset(page);
3482         u64 page_end = start + PAGE_SIZE - 1;
3483         int ret;
3484         int nr = 0;
3485         size_t pg_offset = 0;
3486         loff_t i_size = i_size_read(inode);
3487         unsigned long end_index = i_size >> PAGE_SHIFT;
3488         int write_flags = 0;
3489         unsigned long nr_written = 0;
3490
3491         if (wbc->sync_mode == WB_SYNC_ALL)
3492                 write_flags = REQ_SYNC;
3493
3494         trace___extent_writepage(page, inode, wbc);
3495
3496         WARN_ON(!PageLocked(page));
3497
3498         ClearPageError(page);
3499
3500         pg_offset = i_size & (PAGE_SIZE - 1);
3501         if (page->index > end_index ||
3502            (page->index == end_index && !pg_offset)) {
3503                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3504                 unlock_page(page);
3505                 return 0;
3506         }
3507
3508         if (page->index == end_index) {
3509                 char *userpage;
3510
3511                 userpage = kmap_atomic(page);
3512                 memset(userpage + pg_offset, 0,
3513                        PAGE_SIZE - pg_offset);
3514                 kunmap_atomic(userpage);
3515                 flush_dcache_page(page);
3516         }
3517
3518         pg_offset = 0;
3519
3520         set_page_extent_mapped(page);
3521
3522         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3523         if (ret == 1)
3524                 goto done_unlocked;
3525         if (ret)
3526                 goto done;
3527
3528         ret = __extent_writepage_io(inode, page, wbc, epd,
3529                                     i_size, nr_written, write_flags, &nr);
3530         if (ret == 1)
3531                 goto done_unlocked;
3532
3533 done:
3534         if (nr == 0) {
3535                 /* make sure the mapping tag for page dirty gets cleared */
3536                 set_page_writeback(page);
3537                 end_page_writeback(page);
3538         }
3539         if (PageError(page)) {
3540                 ret = ret < 0 ? ret : -EIO;
3541                 end_extent_writepage(page, ret, start, page_end);
3542         }
3543         unlock_page(page);
3544         return ret;
3545
3546 done_unlocked:
3547         return 0;
3548 }
3549
3550 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3551 {
3552         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3553                        TASK_UNINTERRUPTIBLE);
3554 }
3555
3556 static noinline_for_stack int
3557 lock_extent_buffer_for_io(struct extent_buffer *eb,
3558                           struct btrfs_fs_info *fs_info,
3559                           struct extent_page_data *epd)
3560 {
3561         unsigned long i, num_pages;
3562         int flush = 0;
3563         int ret = 0;
3564
3565         if (!btrfs_try_tree_write_lock(eb)) {
3566                 flush = 1;
3567                 flush_write_bio(epd);
3568                 btrfs_tree_lock(eb);
3569         }
3570
3571         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3572                 btrfs_tree_unlock(eb);
3573                 if (!epd->sync_io)
3574                         return 0;
3575                 if (!flush) {
3576                         flush_write_bio(epd);
3577                         flush = 1;
3578                 }
3579                 while (1) {
3580                         wait_on_extent_buffer_writeback(eb);
3581                         btrfs_tree_lock(eb);
3582                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3583                                 break;
3584                         btrfs_tree_unlock(eb);
3585                 }
3586         }
3587
3588         /*
3589          * We need to do this to prevent races in people who check if the eb is
3590          * under IO since we can end up having no IO bits set for a short period
3591          * of time.
3592          */
3593         spin_lock(&eb->refs_lock);
3594         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3595                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3596                 spin_unlock(&eb->refs_lock);
3597                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3598                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3599                                      -eb->len,
3600                                      fs_info->dirty_metadata_batch);
3601                 ret = 1;
3602         } else {
3603                 spin_unlock(&eb->refs_lock);
3604         }
3605
3606         btrfs_tree_unlock(eb);
3607
3608         if (!ret)
3609                 return ret;
3610
3611         num_pages = num_extent_pages(eb->start, eb->len);
3612         for (i = 0; i < num_pages; i++) {
3613                 struct page *p = eb->pages[i];
3614
3615                 if (!trylock_page(p)) {
3616                         if (!flush) {
3617                                 flush_write_bio(epd);
3618                                 flush = 1;
3619                         }
3620                         lock_page(p);
3621                 }
3622         }
3623
3624         return ret;
3625 }
3626
3627 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3628 {
3629         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3630         smp_mb__after_atomic();
3631         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3632 }
3633
3634 static void set_btree_ioerr(struct page *page)
3635 {
3636         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3637
3638         SetPageError(page);
3639         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3640                 return;
3641
3642         /*
3643          * If writeback for a btree extent that doesn't belong to a log tree
3644          * failed, increment the counter transaction->eb_write_errors.
3645          * We do this because while the transaction is running and before it's
3646          * committing (when we call filemap_fdata[write|wait]_range against
3647          * the btree inode), we might have
3648          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3649          * returns an error or an error happens during writeback, when we're
3650          * committing the transaction we wouldn't know about it, since the pages
3651          * can be no longer dirty nor marked anymore for writeback (if a
3652          * subsequent modification to the extent buffer didn't happen before the
3653          * transaction commit), which makes filemap_fdata[write|wait]_range not
3654          * able to find the pages tagged with SetPageError at transaction
3655          * commit time. So if this happens we must abort the transaction,
3656          * otherwise we commit a super block with btree roots that point to
3657          * btree nodes/leafs whose content on disk is invalid - either garbage
3658          * or the content of some node/leaf from a past generation that got
3659          * cowed or deleted and is no longer valid.
3660          *
3661          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3662          * not be enough - we need to distinguish between log tree extents vs
3663          * non-log tree extents, and the next filemap_fdatawait_range() call
3664          * will catch and clear such errors in the mapping - and that call might
3665          * be from a log sync and not from a transaction commit. Also, checking
3666          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3667          * not done and would not be reliable - the eb might have been released
3668          * from memory and reading it back again means that flag would not be
3669          * set (since it's a runtime flag, not persisted on disk).
3670          *
3671          * Using the flags below in the btree inode also makes us achieve the
3672          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3673          * writeback for all dirty pages and before filemap_fdatawait_range()
3674          * is called, the writeback for all dirty pages had already finished
3675          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3676          * filemap_fdatawait_range() would return success, as it could not know
3677          * that writeback errors happened (the pages were no longer tagged for
3678          * writeback).
3679          */
3680         switch (eb->log_index) {
3681         case -1:
3682                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3683                 break;
3684         case 0:
3685                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3686                 break;
3687         case 1:
3688                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3689                 break;
3690         default:
3691                 BUG(); /* unexpected, logic error */
3692         }
3693 }
3694
3695 static void end_bio_extent_buffer_writepage(struct bio *bio)
3696 {
3697         struct bio_vec *bvec;
3698         struct extent_buffer *eb;
3699         int i, done;
3700
3701         bio_for_each_segment_all(bvec, bio, i) {
3702                 struct page *page = bvec->bv_page;
3703
3704                 eb = (struct extent_buffer *)page->private;
3705                 BUG_ON(!eb);
3706                 done = atomic_dec_and_test(&eb->io_pages);
3707
3708                 if (bio->bi_error ||
3709                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3710                         ClearPageUptodate(page);
3711                         set_btree_ioerr(page);
3712                 }
3713
3714                 end_page_writeback(page);
3715
3716                 if (!done)
3717                         continue;
3718
3719                 end_extent_buffer_writeback(eb);
3720         }
3721
3722         bio_put(bio);
3723 }
3724
3725 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3726                         struct btrfs_fs_info *fs_info,
3727                         struct writeback_control *wbc,
3728                         struct extent_page_data *epd)
3729 {
3730         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3731         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3732         u64 offset = eb->start;
3733         u32 nritems;
3734         unsigned long i, num_pages;
3735         unsigned long bio_flags = 0;
3736         unsigned long start, end;
3737         int write_flags = (epd->sync_io ? REQ_SYNC : 0) | REQ_META;
3738         int ret = 0;
3739
3740         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3741         num_pages = num_extent_pages(eb->start, eb->len);
3742         atomic_set(&eb->io_pages, num_pages);
3743         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3744                 bio_flags = EXTENT_BIO_TREE_LOG;
3745
3746         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3747         nritems = btrfs_header_nritems(eb);
3748         if (btrfs_header_level(eb) > 0) {
3749                 end = btrfs_node_key_ptr_offset(nritems);
3750
3751                 memzero_extent_buffer(eb, end, eb->len - end);
3752         } else {
3753                 /*
3754                  * leaf:
3755                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3756                  */
3757                 start = btrfs_item_nr_offset(nritems);
3758                 end = btrfs_leaf_data(eb) + leaf_data_end(fs_info, eb);
3759                 memzero_extent_buffer(eb, start, end - start);
3760         }
3761
3762         for (i = 0; i < num_pages; i++) {
3763                 struct page *p = eb->pages[i];
3764
3765                 clear_page_dirty_for_io(p);
3766                 set_page_writeback(p);
3767                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3768                                          p, offset >> 9, PAGE_SIZE, 0, bdev,
3769                                          &epd->bio, -1,
3770                                          end_bio_extent_buffer_writepage,
3771                                          0, epd->bio_flags, bio_flags, false);
3772                 epd->bio_flags = bio_flags;
3773                 if (ret) {
3774                         set_btree_ioerr(p);
3775                         end_page_writeback(p);
3776                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3777                                 end_extent_buffer_writeback(eb);
3778                         ret = -EIO;
3779                         break;
3780                 }
3781                 offset += PAGE_SIZE;
3782                 update_nr_written(p, wbc, 1);
3783                 unlock_page(p);
3784         }
3785
3786         if (unlikely(ret)) {
3787                 for (; i < num_pages; i++) {
3788                         struct page *p = eb->pages[i];
3789                         clear_page_dirty_for_io(p);
3790                         unlock_page(p);
3791                 }
3792         }
3793
3794         return ret;
3795 }
3796
3797 int btree_write_cache_pages(struct address_space *mapping,
3798                                    struct writeback_control *wbc)
3799 {
3800         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3801         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3802         struct extent_buffer *eb, *prev_eb = NULL;
3803         struct extent_page_data epd = {
3804                 .bio = NULL,
3805                 .tree = tree,
3806                 .extent_locked = 0,
3807                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3808                 .bio_flags = 0,
3809         };
3810         int ret = 0;
3811         int done = 0;
3812         int nr_to_write_done = 0;
3813         struct pagevec pvec;
3814         int nr_pages;
3815         pgoff_t index;
3816         pgoff_t end;            /* Inclusive */
3817         int scanned = 0;
3818         int tag;
3819
3820         pagevec_init(&pvec, 0);
3821         if (wbc->range_cyclic) {
3822                 index = mapping->writeback_index; /* Start from prev offset */
3823                 end = -1;
3824         } else {
3825                 index = wbc->range_start >> PAGE_SHIFT;
3826                 end = wbc->range_end >> PAGE_SHIFT;
3827                 scanned = 1;
3828         }
3829         if (wbc->sync_mode == WB_SYNC_ALL)
3830                 tag = PAGECACHE_TAG_TOWRITE;
3831         else
3832                 tag = PAGECACHE_TAG_DIRTY;
3833 retry:
3834         if (wbc->sync_mode == WB_SYNC_ALL)
3835                 tag_pages_for_writeback(mapping, index, end);
3836         while (!done && !nr_to_write_done && (index <= end) &&
3837                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3838                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3839                 unsigned i;
3840
3841                 scanned = 1;
3842                 for (i = 0; i < nr_pages; i++) {
3843                         struct page *page = pvec.pages[i];
3844
3845                         if (!PagePrivate(page))
3846                                 continue;
3847
3848                         if (!wbc->range_cyclic && page->index > end) {
3849                                 done = 1;
3850                                 break;
3851                         }
3852
3853                         spin_lock(&mapping->private_lock);
3854                         if (!PagePrivate(page)) {
3855                                 spin_unlock(&mapping->private_lock);
3856                                 continue;
3857                         }
3858
3859                         eb = (struct extent_buffer *)page->private;
3860
3861                         /*
3862                          * Shouldn't happen and normally this would be a BUG_ON
3863                          * but no sense in crashing the users box for something
3864                          * we can survive anyway.
3865                          */
3866                         if (WARN_ON(!eb)) {
3867                                 spin_unlock(&mapping->private_lock);
3868                                 continue;
3869                         }
3870
3871                         if (eb == prev_eb) {
3872                                 spin_unlock(&mapping->private_lock);
3873                                 continue;
3874                         }
3875
3876                         ret = atomic_inc_not_zero(&eb->refs);
3877                         spin_unlock(&mapping->private_lock);
3878                         if (!ret)
3879                                 continue;
3880
3881                         prev_eb = eb;
3882                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3883                         if (!ret) {
3884                                 free_extent_buffer(eb);
3885                                 continue;
3886                         }
3887
3888                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3889                         if (ret) {
3890                                 done = 1;
3891                                 free_extent_buffer(eb);
3892                                 break;
3893                         }
3894                         free_extent_buffer(eb);
3895
3896                         /*
3897                          * the filesystem may choose to bump up nr_to_write.
3898                          * We have to make sure to honor the new nr_to_write
3899                          * at any time
3900                          */
3901                         nr_to_write_done = wbc->nr_to_write <= 0;
3902                 }
3903                 pagevec_release(&pvec);
3904                 cond_resched();
3905         }
3906         if (!scanned && !done) {
3907                 /*
3908                  * We hit the last page and there is more work to be done: wrap
3909                  * back to the start of the file
3910                  */
3911                 scanned = 1;
3912                 index = 0;
3913                 goto retry;
3914         }
3915         flush_write_bio(&epd);
3916         return ret;
3917 }
3918
3919 /**
3920  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3921  * @mapping: address space structure to write
3922  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3923  * @writepage: function called for each page
3924  * @data: data passed to writepage function
3925  *
3926  * If a page is already under I/O, write_cache_pages() skips it, even
3927  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3928  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3929  * and msync() need to guarantee that all the data which was dirty at the time
3930  * the call was made get new I/O started against them.  If wbc->sync_mode is
3931  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3932  * existing IO to complete.
3933  */
3934 static int extent_write_cache_pages(struct extent_io_tree *tree,
3935                              struct address_space *mapping,
3936                              struct writeback_control *wbc,
3937                              writepage_t writepage, void *data,
3938                              void (*flush_fn)(void *))
3939 {
3940         struct inode *inode = mapping->host;
3941         int ret = 0;
3942         int done = 0;
3943         int nr_to_write_done = 0;
3944         struct pagevec pvec;
3945         int nr_pages;
3946         pgoff_t index;
3947         pgoff_t end;            /* Inclusive */
3948         pgoff_t done_index;
3949         int range_whole = 0;
3950         int scanned = 0;
3951         int tag;
3952
3953         /*
3954          * We have to hold onto the inode so that ordered extents can do their
3955          * work when the IO finishes.  The alternative to this is failing to add
3956          * an ordered extent if the igrab() fails there and that is a huge pain
3957          * to deal with, so instead just hold onto the inode throughout the
3958          * writepages operation.  If it fails here we are freeing up the inode
3959          * anyway and we'd rather not waste our time writing out stuff that is
3960          * going to be truncated anyway.
3961          */
3962         if (!igrab(inode))
3963                 return 0;
3964
3965         pagevec_init(&pvec, 0);
3966         if (wbc->range_cyclic) {
3967                 index = mapping->writeback_index; /* Start from prev offset */
3968                 end = -1;
3969         } else {
3970                 index = wbc->range_start >> PAGE_SHIFT;
3971                 end = wbc->range_end >> PAGE_SHIFT;
3972                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3973                         range_whole = 1;
3974                 scanned = 1;
3975         }
3976         if (wbc->sync_mode == WB_SYNC_ALL)
3977                 tag = PAGECACHE_TAG_TOWRITE;
3978         else
3979                 tag = PAGECACHE_TAG_DIRTY;
3980 retry:
3981         if (wbc->sync_mode == WB_SYNC_ALL)
3982                 tag_pages_for_writeback(mapping, index, end);
3983         done_index = index;
3984         while (!done && !nr_to_write_done && (index <= end) &&
3985                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3986                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3987                 unsigned i;
3988
3989                 scanned = 1;
3990                 for (i = 0; i < nr_pages; i++) {
3991                         struct page *page = pvec.pages[i];
3992
3993                         done_index = page->index;
3994                         /*
3995                          * At this point we hold neither mapping->tree_lock nor
3996                          * lock on the page itself: the page may be truncated or
3997                          * invalidated (changing page->mapping to NULL), or even
3998                          * swizzled back from swapper_space to tmpfs file
3999                          * mapping
4000                          */
4001                         if (!trylock_page(page)) {
4002                                 flush_fn(data);
4003                                 lock_page(page);
4004                         }
4005
4006                         if (unlikely(page->mapping != mapping)) {
4007                                 unlock_page(page);
4008                                 continue;
4009                         }
4010
4011                         if (!wbc->range_cyclic && page->index > end) {
4012                                 done = 1;
4013                                 unlock_page(page);
4014                                 continue;
4015                         }
4016
4017                         if (wbc->sync_mode != WB_SYNC_NONE) {
4018                                 if (PageWriteback(page))
4019                                         flush_fn(data);
4020                                 wait_on_page_writeback(page);
4021                         }
4022
4023                         if (PageWriteback(page) ||
4024                             !clear_page_dirty_for_io(page)) {
4025                                 unlock_page(page);
4026                                 continue;
4027                         }
4028
4029                         ret = (*writepage)(page, wbc, data);
4030
4031                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4032                                 unlock_page(page);
4033                                 ret = 0;
4034                         }
4035                         if (ret < 0) {
4036                                 /*
4037                                  * done_index is set past this page,
4038                                  * so media errors will not choke
4039                                  * background writeout for the entire
4040                                  * file. This has consequences for
4041                                  * range_cyclic semantics (ie. it may
4042                                  * not be suitable for data integrity
4043                                  * writeout).
4044                                  */
4045                                 done_index = page->index + 1;
4046                                 done = 1;
4047                                 break;
4048                         }
4049
4050                         /*
4051                          * the filesystem may choose to bump up nr_to_write.
4052                          * We have to make sure to honor the new nr_to_write
4053                          * at any time
4054                          */
4055                         nr_to_write_done = wbc->nr_to_write <= 0;
4056                 }
4057                 pagevec_release(&pvec);
4058                 cond_resched();
4059         }
4060         if (!scanned && !done) {
4061                 /*
4062                  * We hit the last page and there is more work to be done: wrap
4063                  * back to the start of the file
4064                  */
4065                 scanned = 1;
4066                 index = 0;
4067                 goto retry;
4068         }
4069
4070         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4071                 mapping->writeback_index = done_index;
4072
4073         btrfs_add_delayed_iput(inode);
4074         return ret;
4075 }
4076
4077 static void flush_epd_write_bio(struct extent_page_data *epd)
4078 {
4079         if (epd->bio) {
4080                 int ret;
4081
4082                 bio_set_op_attrs(epd->bio, REQ_OP_WRITE,
4083                                  epd->sync_io ? REQ_SYNC : 0);
4084
4085                 ret = submit_one_bio(epd->bio, 0, epd->bio_flags);
4086                 BUG_ON(ret < 0); /* -ENOMEM */
4087                 epd->bio = NULL;
4088         }
4089 }
4090
4091 static noinline void flush_write_bio(void *data)
4092 {
4093         struct extent_page_data *epd = data;
4094         flush_epd_write_bio(epd);
4095 }
4096
4097 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4098                           get_extent_t *get_extent,
4099                           struct writeback_control *wbc)
4100 {
4101         int ret;
4102         struct extent_page_data epd = {
4103                 .bio = NULL,
4104                 .tree = tree,
4105                 .get_extent = get_extent,
4106                 .extent_locked = 0,
4107                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4108                 .bio_flags = 0,
4109         };
4110
4111         ret = __extent_writepage(page, wbc, &epd);
4112
4113         flush_epd_write_bio(&epd);
4114         return ret;
4115 }
4116
4117 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4118                               u64 start, u64 end, get_extent_t *get_extent,
4119                               int mode)
4120 {
4121         int ret = 0;
4122         struct address_space *mapping = inode->i_mapping;
4123         struct page *page;
4124         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4125                 PAGE_SHIFT;
4126
4127         struct extent_page_data epd = {
4128                 .bio = NULL,
4129                 .tree = tree,
4130                 .get_extent = get_extent,
4131                 .extent_locked = 1,
4132                 .sync_io = mode == WB_SYNC_ALL,
4133                 .bio_flags = 0,
4134         };
4135         struct writeback_control wbc_writepages = {
4136                 .sync_mode      = mode,
4137                 .nr_to_write    = nr_pages * 2,
4138                 .range_start    = start,
4139                 .range_end      = end + 1,
4140         };
4141
4142         while (start <= end) {
4143                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4144                 if (clear_page_dirty_for_io(page))
4145                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4146                 else {
4147                         if (tree->ops && tree->ops->writepage_end_io_hook)
4148                                 tree->ops->writepage_end_io_hook(page, start,
4149                                                  start + PAGE_SIZE - 1,
4150                                                  NULL, 1);
4151                         unlock_page(page);
4152                 }
4153                 put_page(page);
4154                 start += PAGE_SIZE;
4155         }
4156
4157         flush_epd_write_bio(&epd);
4158         return ret;
4159 }
4160
4161 int extent_writepages(struct extent_io_tree *tree,
4162                       struct address_space *mapping,
4163                       get_extent_t *get_extent,
4164                       struct writeback_control *wbc)
4165 {
4166         int ret = 0;
4167         struct extent_page_data epd = {
4168                 .bio = NULL,
4169                 .tree = tree,
4170                 .get_extent = get_extent,
4171                 .extent_locked = 0,
4172                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4173                 .bio_flags = 0,
4174         };
4175
4176         ret = extent_write_cache_pages(tree, mapping, wbc,
4177                                        __extent_writepage, &epd,
4178                                        flush_write_bio);
4179         flush_epd_write_bio(&epd);
4180         return ret;
4181 }
4182
4183 int extent_readpages(struct extent_io_tree *tree,
4184                      struct address_space *mapping,
4185                      struct list_head *pages, unsigned nr_pages,
4186                      get_extent_t get_extent)
4187 {
4188         struct bio *bio = NULL;
4189         unsigned page_idx;
4190         unsigned long bio_flags = 0;
4191         struct page *pagepool[16];
4192         struct page *page;
4193         struct extent_map *em_cached = NULL;
4194         int nr = 0;
4195         u64 prev_em_start = (u64)-1;
4196
4197         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4198                 page = list_entry(pages->prev, struct page, lru);
4199
4200                 prefetchw(&page->flags);
4201                 list_del(&page->lru);
4202                 if (add_to_page_cache_lru(page, mapping,
4203                                         page->index,
4204                                         readahead_gfp_mask(mapping))) {
4205                         put_page(page);
4206                         continue;
4207                 }
4208
4209                 pagepool[nr++] = page;
4210                 if (nr < ARRAY_SIZE(pagepool))
4211                         continue;
4212                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4213                                    &bio, 0, &bio_flags, &prev_em_start);
4214                 nr = 0;
4215         }
4216         if (nr)
4217                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4218                                    &bio, 0, &bio_flags, &prev_em_start);
4219
4220         if (em_cached)
4221                 free_extent_map(em_cached);
4222
4223         BUG_ON(!list_empty(pages));
4224         if (bio)
4225                 return submit_one_bio(bio, 0, bio_flags);
4226         return 0;
4227 }
4228
4229 /*
4230  * basic invalidatepage code, this waits on any locked or writeback
4231  * ranges corresponding to the page, and then deletes any extent state
4232  * records from the tree
4233  */
4234 int extent_invalidatepage(struct extent_io_tree *tree,
4235                           struct page *page, unsigned long offset)
4236 {
4237         struct extent_state *cached_state = NULL;
4238         u64 start = page_offset(page);
4239         u64 end = start + PAGE_SIZE - 1;
4240         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4241
4242         start += ALIGN(offset, blocksize);
4243         if (start > end)
4244                 return 0;
4245
4246         lock_extent_bits(tree, start, end, &cached_state);
4247         wait_on_page_writeback(page);
4248         clear_extent_bit(tree, start, end,
4249                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4250                          EXTENT_DO_ACCOUNTING,
4251                          1, 1, &cached_state, GFP_NOFS);
4252         return 0;
4253 }
4254
4255 /*
4256  * a helper for releasepage, this tests for areas of the page that
4257  * are locked or under IO and drops the related state bits if it is safe
4258  * to drop the page.
4259  */
4260 static int try_release_extent_state(struct extent_map_tree *map,
4261                                     struct extent_io_tree *tree,
4262                                     struct page *page, gfp_t mask)
4263 {
4264         u64 start = page_offset(page);
4265         u64 end = start + PAGE_SIZE - 1;
4266         int ret = 1;
4267
4268         if (test_range_bit(tree, start, end,
4269                            EXTENT_IOBITS, 0, NULL))
4270                 ret = 0;
4271         else {
4272                 /*
4273                  * at this point we can safely clear everything except the
4274                  * locked bit and the nodatasum bit
4275                  */
4276                 ret = clear_extent_bit(tree, start, end,
4277                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4278                                  0, 0, NULL, mask);
4279
4280                 /* if clear_extent_bit failed for enomem reasons,
4281                  * we can't allow the release to continue.
4282                  */
4283                 if (ret < 0)
4284                         ret = 0;
4285                 else
4286                         ret = 1;
4287         }
4288         return ret;
4289 }
4290
4291 /*
4292  * a helper for releasepage.  As long as there are no locked extents
4293  * in the range corresponding to the page, both state records and extent
4294  * map records are removed
4295  */
4296 int try_release_extent_mapping(struct extent_map_tree *map,
4297                                struct extent_io_tree *tree, struct page *page,
4298                                gfp_t mask)
4299 {
4300         struct extent_map *em;
4301         u64 start = page_offset(page);
4302         u64 end = start + PAGE_SIZE - 1;
4303
4304         if (gfpflags_allow_blocking(mask) &&
4305             page->mapping->host->i_size > SZ_16M) {
4306                 u64 len;
4307                 while (start <= end) {
4308                         len = end - start + 1;
4309                         write_lock(&map->lock);
4310                         em = lookup_extent_mapping(map, start, len);
4311                         if (!em) {
4312                                 write_unlock(&map->lock);
4313                                 break;
4314                         }
4315                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4316                             em->start != start) {
4317                                 write_unlock(&map->lock);
4318                                 free_extent_map(em);
4319                                 break;
4320                         }
4321                         if (!test_range_bit(tree, em->start,
4322                                             extent_map_end(em) - 1,
4323                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4324                                             0, NULL)) {
4325                                 remove_extent_mapping(map, em);
4326                                 /* once for the rb tree */
4327                                 free_extent_map(em);
4328                         }
4329                         start = extent_map_end(em);
4330                         write_unlock(&map->lock);
4331
4332                         /* once for us */
4333                         free_extent_map(em);
4334                 }
4335         }
4336         return try_release_extent_state(map, tree, page, mask);
4337 }
4338
4339 /*
4340  * helper function for fiemap, which doesn't want to see any holes.
4341  * This maps until we find something past 'last'
4342  */
4343 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4344                                                 u64 offset,
4345                                                 u64 last,
4346                                                 get_extent_t *get_extent)
4347 {
4348         u64 sectorsize = btrfs_inode_sectorsize(inode);
4349         struct extent_map *em;
4350         u64 len;
4351
4352         if (offset >= last)
4353                 return NULL;
4354
4355         while (1) {
4356                 len = last - offset;
4357                 if (len == 0)
4358                         break;
4359                 len = ALIGN(len, sectorsize);
4360                 em = get_extent(inode, NULL, 0, offset, len, 0);
4361                 if (IS_ERR_OR_NULL(em))
4362                         return em;
4363
4364                 /* if this isn't a hole return it */
4365                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4366                     em->block_start != EXTENT_MAP_HOLE) {
4367                         return em;
4368                 }
4369
4370                 /* this is a hole, advance to the next extent */
4371                 offset = extent_map_end(em);
4372                 free_extent_map(em);
4373                 if (offset >= last)
4374                         break;
4375         }
4376         return NULL;
4377 }
4378
4379 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4380                 __u64 start, __u64 len, get_extent_t *get_extent)
4381 {
4382         int ret = 0;
4383         u64 off = start;
4384         u64 max = start + len;
4385         u32 flags = 0;
4386         u32 found_type;
4387         u64 last;
4388         u64 last_for_get_extent = 0;
4389         u64 disko = 0;
4390         u64 isize = i_size_read(inode);
4391         struct btrfs_key found_key;
4392         struct extent_map *em = NULL;
4393         struct extent_state *cached_state = NULL;
4394         struct btrfs_path *path;
4395         struct btrfs_root *root = BTRFS_I(inode)->root;
4396         int end = 0;
4397         u64 em_start = 0;
4398         u64 em_len = 0;
4399         u64 em_end = 0;
4400
4401         if (len == 0)
4402                 return -EINVAL;
4403
4404         path = btrfs_alloc_path();
4405         if (!path)
4406                 return -ENOMEM;
4407         path->leave_spinning = 1;
4408
4409         start = round_down(start, btrfs_inode_sectorsize(inode));
4410         len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
4411
4412         /*
4413          * lookup the last file extent.  We're not using i_size here
4414          * because there might be preallocation past i_size
4415          */
4416         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(BTRFS_I(inode)), -1,
4417                                        0);
4418         if (ret < 0) {
4419                 btrfs_free_path(path);
4420                 return ret;
4421         } else {
4422                 WARN_ON(!ret);
4423                 if (ret == 1)
4424                         ret = 0;
4425         }
4426
4427         path->slots[0]--;
4428         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4429         found_type = found_key.type;
4430
4431         /* No extents, but there might be delalloc bits */
4432         if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) ||
4433             found_type != BTRFS_EXTENT_DATA_KEY) {
4434                 /* have to trust i_size as the end */
4435                 last = (u64)-1;
4436                 last_for_get_extent = isize;
4437         } else {
4438                 /*
4439                  * remember the start of the last extent.  There are a
4440                  * bunch of different factors that go into the length of the
4441                  * extent, so its much less complex to remember where it started
4442                  */
4443                 last = found_key.offset;
4444                 last_for_get_extent = last + 1;
4445         }
4446         btrfs_release_path(path);
4447
4448         /*
4449          * we might have some extents allocated but more delalloc past those
4450          * extents.  so, we trust isize unless the start of the last extent is
4451          * beyond isize
4452          */
4453         if (last < isize) {
4454                 last = (u64)-1;
4455                 last_for_get_extent = isize;
4456         }
4457
4458         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4459                          &cached_state);
4460
4461         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4462                                    get_extent);
4463         if (!em)
4464                 goto out;
4465         if (IS_ERR(em)) {
4466                 ret = PTR_ERR(em);
4467                 goto out;
4468         }
4469
4470         while (!end) {
4471                 u64 offset_in_extent = 0;
4472
4473                 /* break if the extent we found is outside the range */
4474                 if (em->start >= max || extent_map_end(em) < off)
4475                         break;
4476
4477                 /*
4478                  * get_extent may return an extent that starts before our
4479                  * requested range.  We have to make sure the ranges
4480                  * we return to fiemap always move forward and don't
4481                  * overlap, so adjust the offsets here
4482                  */
4483                 em_start = max(em->start, off);
4484
4485                 /*
4486                  * record the offset from the start of the extent
4487                  * for adjusting the disk offset below.  Only do this if the
4488                  * extent isn't compressed since our in ram offset may be past
4489                  * what we have actually allocated on disk.
4490                  */
4491                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4492                         offset_in_extent = em_start - em->start;
4493                 em_end = extent_map_end(em);
4494                 em_len = em_end - em_start;
4495                 disko = 0;
4496                 flags = 0;
4497
4498                 /*
4499                  * bump off for our next call to get_extent
4500                  */
4501                 off = extent_map_end(em);
4502                 if (off >= max)
4503                         end = 1;
4504
4505                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4506                         end = 1;
4507                         flags |= FIEMAP_EXTENT_LAST;
4508                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4509                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4510                                   FIEMAP_EXTENT_NOT_ALIGNED);
4511                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4512                         flags |= (FIEMAP_EXTENT_DELALLOC |
4513                                   FIEMAP_EXTENT_UNKNOWN);
4514                 } else if (fieinfo->fi_extents_max) {
4515                         struct btrfs_trans_handle *trans;
4516
4517                         u64 bytenr = em->block_start -
4518                                 (em->start - em->orig_start);
4519
4520                         disko = em->block_start + offset_in_extent;
4521
4522                         /*
4523                          * We need a trans handle to get delayed refs
4524                          */
4525                         trans = btrfs_join_transaction(root);
4526                         /*
4527                          * It's OK if we can't start a trans we can still check
4528                          * from commit_root
4529                          */
4530                         if (IS_ERR(trans))
4531                                 trans = NULL;
4532
4533                         /*
4534                          * As btrfs supports shared space, this information
4535                          * can be exported to userspace tools via
4536                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4537                          * then we're just getting a count and we can skip the
4538                          * lookup stuff.
4539                          */
4540                         ret = btrfs_check_shared(trans, root->fs_info,
4541                                                  root->objectid,
4542                                                  btrfs_ino(BTRFS_I(inode)), bytenr);
4543                         if (trans)
4544                                 btrfs_end_transaction(trans);
4545                         if (ret < 0)
4546                                 goto out_free;
4547                         if (ret)
4548                                 flags |= FIEMAP_EXTENT_SHARED;
4549                         ret = 0;
4550                 }
4551                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4552                         flags |= FIEMAP_EXTENT_ENCODED;
4553                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4554                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4555
4556                 free_extent_map(em);
4557                 em = NULL;
4558                 if ((em_start >= last) || em_len == (u64)-1 ||
4559                    (last == (u64)-1 && isize <= em_end)) {
4560                         flags |= FIEMAP_EXTENT_LAST;
4561                         end = 1;
4562                 }
4563
4564                 /* now scan forward to see if this is really the last extent. */
4565                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4566                                            get_extent);
4567                 if (IS_ERR(em)) {
4568                         ret = PTR_ERR(em);
4569                         goto out;
4570                 }
4571                 if (!em) {
4572                         flags |= FIEMAP_EXTENT_LAST;
4573                         end = 1;
4574                 }
4575                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4576                                               em_len, flags);
4577                 if (ret) {
4578                         if (ret == 1)
4579                                 ret = 0;
4580                         goto out_free;
4581                 }
4582         }
4583 out_free:
4584         free_extent_map(em);
4585 out:
4586         btrfs_free_path(path);
4587         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4588                              &cached_state, GFP_NOFS);
4589         return ret;
4590 }
4591
4592 static void __free_extent_buffer(struct extent_buffer *eb)
4593 {
4594         btrfs_leak_debug_del(&eb->leak_list);
4595         kmem_cache_free(extent_buffer_cache, eb);
4596 }
4597
4598 int extent_buffer_under_io(struct extent_buffer *eb)
4599 {
4600         return (atomic_read(&eb->io_pages) ||
4601                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4602                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4603 }
4604
4605 /*
4606  * Helper for releasing extent buffer page.
4607  */
4608 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4609 {
4610         unsigned long index;
4611         struct page *page;
4612         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4613
4614         BUG_ON(extent_buffer_under_io(eb));
4615
4616         index = num_extent_pages(eb->start, eb->len);
4617         if (index == 0)
4618                 return;
4619
4620         do {
4621                 index--;
4622                 page = eb->pages[index];
4623                 if (!page)
4624                         continue;
4625                 if (mapped)
4626                         spin_lock(&page->mapping->private_lock);
4627                 /*
4628                  * We do this since we'll remove the pages after we've
4629                  * removed the eb from the radix tree, so we could race
4630                  * and have this page now attached to the new eb.  So
4631                  * only clear page_private if it's still connected to
4632                  * this eb.
4633                  */
4634                 if (PagePrivate(page) &&
4635                     page->private == (unsigned long)eb) {
4636                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4637                         BUG_ON(PageDirty(page));
4638                         BUG_ON(PageWriteback(page));
4639                         /*
4640                          * We need to make sure we haven't be attached
4641                          * to a new eb.
4642                          */
4643                         ClearPagePrivate(page);
4644                         set_page_private(page, 0);
4645                         /* One for the page private */
4646                         put_page(page);
4647                 }
4648
4649                 if (mapped)
4650                         spin_unlock(&page->mapping->private_lock);
4651
4652                 /* One for when we allocated the page */
4653                 put_page(page);
4654         } while (index != 0);
4655 }
4656
4657 /*
4658  * Helper for releasing the extent buffer.
4659  */
4660 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4661 {
4662         btrfs_release_extent_buffer_page(eb);
4663         __free_extent_buffer(eb);
4664 }
4665
4666 static struct extent_buffer *
4667 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4668                       unsigned long len)
4669 {
4670         struct extent_buffer *eb = NULL;
4671
4672         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4673         eb->start = start;
4674         eb->len = len;
4675         eb->fs_info = fs_info;
4676         eb->bflags = 0;
4677         rwlock_init(&eb->lock);
4678         atomic_set(&eb->write_locks, 0);
4679         atomic_set(&eb->read_locks, 0);
4680         atomic_set(&eb->blocking_readers, 0);
4681         atomic_set(&eb->blocking_writers, 0);
4682         atomic_set(&eb->spinning_readers, 0);
4683         atomic_set(&eb->spinning_writers, 0);
4684         eb->lock_nested = 0;
4685         init_waitqueue_head(&eb->write_lock_wq);
4686         init_waitqueue_head(&eb->read_lock_wq);
4687
4688         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4689
4690         spin_lock_init(&eb->refs_lock);
4691         atomic_set(&eb->refs, 1);
4692         atomic_set(&eb->io_pages, 0);
4693
4694         /*
4695          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4696          */
4697         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4698                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4699         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4700
4701         return eb;
4702 }
4703
4704 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4705 {
4706         unsigned long i;
4707         struct page *p;
4708         struct extent_buffer *new;
4709         unsigned long num_pages = num_extent_pages(src->start, src->len);
4710
4711         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4712         if (new == NULL)
4713                 return NULL;
4714
4715         for (i = 0; i < num_pages; i++) {
4716                 p = alloc_page(GFP_NOFS);
4717                 if (!p) {
4718                         btrfs_release_extent_buffer(new);
4719                         return NULL;
4720                 }
4721                 attach_extent_buffer_page(new, p);
4722                 WARN_ON(PageDirty(p));
4723                 SetPageUptodate(p);
4724                 new->pages[i] = p;
4725                 copy_page(page_address(p), page_address(src->pages[i]));
4726         }
4727
4728         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4729         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4730
4731         return new;
4732 }
4733
4734 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4735                                                   u64 start, unsigned long len)
4736 {
4737         struct extent_buffer *eb;
4738         unsigned long num_pages;
4739         unsigned long i;
4740
4741         num_pages = num_extent_pages(start, len);
4742
4743         eb = __alloc_extent_buffer(fs_info, start, len);
4744         if (!eb)
4745                 return NULL;
4746
4747         for (i = 0; i < num_pages; i++) {
4748                 eb->pages[i] = alloc_page(GFP_NOFS);
4749                 if (!eb->pages[i])
4750                         goto err;
4751         }
4752         set_extent_buffer_uptodate(eb);
4753         btrfs_set_header_nritems(eb, 0);
4754         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4755
4756         return eb;
4757 err:
4758         for (; i > 0; i--)
4759                 __free_page(eb->pages[i - 1]);
4760         __free_extent_buffer(eb);
4761         return NULL;
4762 }
4763
4764 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4765                                                 u64 start)
4766 {
4767         return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
4768 }
4769
4770 static void check_buffer_tree_ref(struct extent_buffer *eb)
4771 {
4772         int refs;
4773         /* the ref bit is tricky.  We have to make sure it is set
4774          * if we have the buffer dirty.   Otherwise the
4775          * code to free a buffer can end up dropping a dirty
4776          * page
4777          *
4778          * Once the ref bit is set, it won't go away while the
4779          * buffer is dirty or in writeback, and it also won't
4780          * go away while we have the reference count on the
4781          * eb bumped.
4782          *
4783          * We can't just set the ref bit without bumping the
4784          * ref on the eb because free_extent_buffer might
4785          * see the ref bit and try to clear it.  If this happens
4786          * free_extent_buffer might end up dropping our original
4787          * ref by mistake and freeing the page before we are able
4788          * to add one more ref.
4789          *
4790          * So bump the ref count first, then set the bit.  If someone
4791          * beat us to it, drop the ref we added.
4792          */
4793         refs = atomic_read(&eb->refs);
4794         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4795                 return;
4796
4797         spin_lock(&eb->refs_lock);
4798         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4799                 atomic_inc(&eb->refs);
4800         spin_unlock(&eb->refs_lock);
4801 }
4802
4803 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4804                 struct page *accessed)
4805 {
4806         unsigned long num_pages, i;
4807
4808         check_buffer_tree_ref(eb);
4809
4810         num_pages = num_extent_pages(eb->start, eb->len);
4811         for (i = 0; i < num_pages; i++) {
4812                 struct page *p = eb->pages[i];
4813
4814                 if (p != accessed)
4815                         mark_page_accessed(p);
4816         }
4817 }
4818
4819 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4820                                          u64 start)
4821 {
4822         struct extent_buffer *eb;
4823
4824         rcu_read_lock();
4825         eb = radix_tree_lookup(&fs_info->buffer_radix,
4826                                start >> PAGE_SHIFT);
4827         if (eb && atomic_inc_not_zero(&eb->refs)) {
4828                 rcu_read_unlock();
4829                 /*
4830                  * Lock our eb's refs_lock to avoid races with
4831                  * free_extent_buffer. When we get our eb it might be flagged
4832                  * with EXTENT_BUFFER_STALE and another task running
4833                  * free_extent_buffer might have seen that flag set,
4834                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4835                  * writeback flags not set) and it's still in the tree (flag
4836                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4837                  * of decrementing the extent buffer's reference count twice.
4838                  * So here we could race and increment the eb's reference count,
4839                  * clear its stale flag, mark it as dirty and drop our reference
4840                  * before the other task finishes executing free_extent_buffer,
4841                  * which would later result in an attempt to free an extent
4842                  * buffer that is dirty.
4843                  */
4844                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4845                         spin_lock(&eb->refs_lock);
4846                         spin_unlock(&eb->refs_lock);
4847                 }
4848                 mark_extent_buffer_accessed(eb, NULL);
4849                 return eb;
4850         }
4851         rcu_read_unlock();
4852
4853         return NULL;
4854 }
4855
4856 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4857 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4858                                         u64 start)
4859 {
4860         struct extent_buffer *eb, *exists = NULL;
4861         int ret;
4862
4863         eb = find_extent_buffer(fs_info, start);
4864         if (eb)
4865                 return eb;
4866         eb = alloc_dummy_extent_buffer(fs_info, start);
4867         if (!eb)
4868                 return NULL;
4869         eb->fs_info = fs_info;
4870 again:
4871         ret = radix_tree_preload(GFP_NOFS);
4872         if (ret)
4873                 goto free_eb;
4874         spin_lock(&fs_info->buffer_lock);
4875         ret = radix_tree_insert(&fs_info->buffer_radix,
4876                                 start >> PAGE_SHIFT, eb);
4877         spin_unlock(&fs_info->buffer_lock);
4878         radix_tree_preload_end();
4879         if (ret == -EEXIST) {
4880                 exists = find_extent_buffer(fs_info, start);
4881                 if (exists)
4882                         goto free_eb;
4883                 else
4884                         goto again;
4885         }
4886         check_buffer_tree_ref(eb);
4887         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4888
4889         /*
4890          * We will free dummy extent buffer's if they come into
4891          * free_extent_buffer with a ref count of 2, but if we are using this we
4892          * want the buffers to stay in memory until we're done with them, so
4893          * bump the ref count again.
4894          */
4895         atomic_inc(&eb->refs);
4896         return eb;
4897 free_eb:
4898         btrfs_release_extent_buffer(eb);
4899         return exists;
4900 }
4901 #endif
4902
4903 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4904                                           u64 start)
4905 {
4906         unsigned long len = fs_info->nodesize;
4907         unsigned long num_pages = num_extent_pages(start, len);
4908         unsigned long i;
4909         unsigned long index = start >> PAGE_SHIFT;
4910         struct extent_buffer *eb;
4911         struct extent_buffer *exists = NULL;
4912         struct page *p;
4913         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4914         int uptodate = 1;
4915         int ret;
4916
4917         if (!IS_ALIGNED(start, fs_info->sectorsize)) {
4918                 btrfs_err(fs_info, "bad tree block start %llu", start);
4919                 return ERR_PTR(-EINVAL);
4920         }
4921
4922         eb = find_extent_buffer(fs_info, start);
4923         if (eb)
4924                 return eb;
4925
4926         eb = __alloc_extent_buffer(fs_info, start, len);
4927         if (!eb)
4928                 return ERR_PTR(-ENOMEM);
4929
4930         for (i = 0; i < num_pages; i++, index++) {
4931                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4932                 if (!p) {
4933                         exists = ERR_PTR(-ENOMEM);
4934                         goto free_eb;
4935                 }
4936
4937                 spin_lock(&mapping->private_lock);
4938                 if (PagePrivate(p)) {
4939                         /*
4940                          * We could have already allocated an eb for this page
4941                          * and attached one so lets see if we can get a ref on
4942                          * the existing eb, and if we can we know it's good and
4943                          * we can just return that one, else we know we can just
4944                          * overwrite page->private.
4945                          */
4946                         exists = (struct extent_buffer *)p->private;
4947                         if (atomic_inc_not_zero(&exists->refs)) {
4948                                 spin_unlock(&mapping->private_lock);
4949                                 unlock_page(p);
4950                                 put_page(p);
4951                                 mark_extent_buffer_accessed(exists, p);
4952                                 goto free_eb;
4953                         }
4954                         exists = NULL;
4955
4956                         /*
4957                          * Do this so attach doesn't complain and we need to
4958                          * drop the ref the old guy had.
4959                          */
4960                         ClearPagePrivate(p);
4961                         WARN_ON(PageDirty(p));
4962                         put_page(p);
4963                 }
4964                 attach_extent_buffer_page(eb, p);
4965                 spin_unlock(&mapping->private_lock);
4966                 WARN_ON(PageDirty(p));
4967                 eb->pages[i] = p;
4968                 if (!PageUptodate(p))
4969                         uptodate = 0;
4970
4971                 /*
4972                  * see below about how we avoid a nasty race with release page
4973                  * and why we unlock later
4974                  */
4975         }
4976         if (uptodate)
4977                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4978 again:
4979         ret = radix_tree_preload(GFP_NOFS);
4980         if (ret) {
4981                 exists = ERR_PTR(ret);
4982                 goto free_eb;
4983         }
4984
4985         spin_lock(&fs_info->buffer_lock);
4986         ret = radix_tree_insert(&fs_info->buffer_radix,
4987                                 start >> PAGE_SHIFT, eb);
4988         spin_unlock(&fs_info->buffer_lock);
4989         radix_tree_preload_end();
4990         if (ret == -EEXIST) {
4991                 exists = find_extent_buffer(fs_info, start);
4992                 if (exists)
4993                         goto free_eb;
4994                 else
4995                         goto again;
4996         }
4997         /* add one reference for the tree */
4998         check_buffer_tree_ref(eb);
4999         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5000
5001         /*
5002          * there is a race where release page may have
5003          * tried to find this extent buffer in the radix
5004          * but failed.  It will tell the VM it is safe to
5005          * reclaim the, and it will clear the page private bit.
5006          * We must make sure to set the page private bit properly
5007          * after the extent buffer is in the radix tree so
5008          * it doesn't get lost
5009          */
5010         SetPageChecked(eb->pages[0]);
5011         for (i = 1; i < num_pages; i++) {
5012                 p = eb->pages[i];
5013                 ClearPageChecked(p);
5014                 unlock_page(p);
5015         }
5016         unlock_page(eb->pages[0]);
5017         return eb;
5018
5019 free_eb:
5020         WARN_ON(!atomic_dec_and_test(&eb->refs));
5021         for (i = 0; i < num_pages; i++) {
5022                 if (eb->pages[i])
5023                         unlock_page(eb->pages[i]);
5024         }
5025
5026         btrfs_release_extent_buffer(eb);
5027         return exists;
5028 }
5029
5030 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5031 {
5032         struct extent_buffer *eb =
5033                         container_of(head, struct extent_buffer, rcu_head);
5034
5035         __free_extent_buffer(eb);
5036 }
5037
5038 /* Expects to have eb->eb_lock already held */
5039 static int release_extent_buffer(struct extent_buffer *eb)
5040 {
5041         WARN_ON(atomic_read(&eb->refs) == 0);
5042         if (atomic_dec_and_test(&eb->refs)) {
5043                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5044                         struct btrfs_fs_info *fs_info = eb->fs_info;
5045
5046                         spin_unlock(&eb->refs_lock);
5047
5048                         spin_lock(&fs_info->buffer_lock);
5049                         radix_tree_delete(&fs_info->buffer_radix,
5050                                           eb->start >> PAGE_SHIFT);
5051                         spin_unlock(&fs_info->buffer_lock);
5052                 } else {
5053                         spin_unlock(&eb->refs_lock);
5054                 }
5055
5056                 /* Should be safe to release our pages at this point */
5057                 btrfs_release_extent_buffer_page(eb);
5058 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5059                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5060                         __free_extent_buffer(eb);
5061                         return 1;
5062                 }
5063 #endif
5064                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5065                 return 1;
5066         }
5067         spin_unlock(&eb->refs_lock);
5068
5069         return 0;
5070 }
5071
5072 void free_extent_buffer(struct extent_buffer *eb)
5073 {
5074         int refs;
5075         int old;
5076         if (!eb)
5077                 return;
5078
5079         while (1) {
5080                 refs = atomic_read(&eb->refs);
5081                 if (refs <= 3)
5082                         break;
5083                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5084                 if (old == refs)
5085                         return;
5086         }
5087
5088         spin_lock(&eb->refs_lock);
5089         if (atomic_read(&eb->refs) == 2 &&
5090             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5091                 atomic_dec(&eb->refs);
5092
5093         if (atomic_read(&eb->refs) == 2 &&
5094             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5095             !extent_buffer_under_io(eb) &&
5096             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5097                 atomic_dec(&eb->refs);
5098
5099         /*
5100          * I know this is terrible, but it's temporary until we stop tracking
5101          * the uptodate bits and such for the extent buffers.
5102          */
5103         release_extent_buffer(eb);
5104 }
5105
5106 void free_extent_buffer_stale(struct extent_buffer *eb)
5107 {
5108         if (!eb)
5109                 return;
5110
5111         spin_lock(&eb->refs_lock);
5112         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5113
5114         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5115             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5116                 atomic_dec(&eb->refs);
5117         release_extent_buffer(eb);
5118 }
5119
5120 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5121 {
5122         unsigned long i;
5123         unsigned long num_pages;
5124         struct page *page;
5125
5126         num_pages = num_extent_pages(eb->start, eb->len);
5127
5128         for (i = 0; i < num_pages; i++) {
5129                 page = eb->pages[i];
5130                 if (!PageDirty(page))
5131                         continue;
5132
5133                 lock_page(page);
5134                 WARN_ON(!PagePrivate(page));
5135
5136                 clear_page_dirty_for_io(page);
5137                 spin_lock_irq(&page->mapping->tree_lock);
5138                 if (!PageDirty(page)) {
5139                         radix_tree_tag_clear(&page->mapping->page_tree,
5140                                                 page_index(page),
5141                                                 PAGECACHE_TAG_DIRTY);
5142                 }
5143                 spin_unlock_irq(&page->mapping->tree_lock);
5144                 ClearPageError(page);
5145                 unlock_page(page);
5146         }
5147         WARN_ON(atomic_read(&eb->refs) == 0);
5148 }
5149
5150 int set_extent_buffer_dirty(struct extent_buffer *eb)
5151 {
5152         unsigned long i;
5153         unsigned long num_pages;
5154         int was_dirty = 0;
5155
5156         check_buffer_tree_ref(eb);
5157
5158         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5159
5160         num_pages = num_extent_pages(eb->start, eb->len);
5161         WARN_ON(atomic_read(&eb->refs) == 0);
5162         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5163
5164         for (i = 0; i < num_pages; i++)
5165                 set_page_dirty(eb->pages[i]);
5166         return was_dirty;
5167 }
5168
5169 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5170 {
5171         unsigned long i;
5172         struct page *page;
5173         unsigned long num_pages;
5174
5175         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5176         num_pages = num_extent_pages(eb->start, eb->len);
5177         for (i = 0; i < num_pages; i++) {
5178                 page = eb->pages[i];
5179                 if (page)
5180                         ClearPageUptodate(page);
5181         }
5182 }
5183
5184 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5185 {
5186         unsigned long i;
5187         struct page *page;
5188         unsigned long num_pages;
5189
5190         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5191         num_pages = num_extent_pages(eb->start, eb->len);
5192         for (i = 0; i < num_pages; i++) {
5193                 page = eb->pages[i];
5194                 SetPageUptodate(page);
5195         }
5196 }
5197
5198 int extent_buffer_uptodate(struct extent_buffer *eb)
5199 {
5200         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5201 }
5202
5203 int read_extent_buffer_pages(struct extent_io_tree *tree,
5204                              struct extent_buffer *eb, int wait,
5205                              get_extent_t *get_extent, int mirror_num)
5206 {
5207         unsigned long i;
5208         struct page *page;
5209         int err;
5210         int ret = 0;
5211         int locked_pages = 0;
5212         int all_uptodate = 1;
5213         unsigned long num_pages;
5214         unsigned long num_reads = 0;
5215         struct bio *bio = NULL;
5216         unsigned long bio_flags = 0;
5217
5218         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5219                 return 0;
5220
5221         num_pages = num_extent_pages(eb->start, eb->len);
5222         for (i = 0; i < num_pages; i++) {
5223                 page = eb->pages[i];
5224                 if (wait == WAIT_NONE) {
5225                         if (!trylock_page(page))
5226                                 goto unlock_exit;
5227                 } else {
5228                         lock_page(page);
5229                 }
5230                 locked_pages++;
5231         }
5232         /*
5233          * We need to firstly lock all pages to make sure that
5234          * the uptodate bit of our pages won't be affected by
5235          * clear_extent_buffer_uptodate().
5236          */
5237         for (i = 0; i < num_pages; i++) {
5238                 page = eb->pages[i];
5239                 if (!PageUptodate(page)) {
5240                         num_reads++;
5241                         all_uptodate = 0;
5242                 }
5243         }
5244
5245         if (all_uptodate) {
5246                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5247                 goto unlock_exit;
5248         }
5249
5250         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5251         eb->read_mirror = 0;
5252         atomic_set(&eb->io_pages, num_reads);
5253         for (i = 0; i < num_pages; i++) {
5254                 page = eb->pages[i];
5255
5256                 if (!PageUptodate(page)) {
5257                         if (ret) {
5258                                 atomic_dec(&eb->io_pages);
5259                                 unlock_page(page);
5260                                 continue;
5261                         }
5262
5263                         ClearPageError(page);
5264                         err = __extent_read_full_page(tree, page,
5265                                                       get_extent, &bio,
5266                                                       mirror_num, &bio_flags,
5267                                                       REQ_META);
5268                         if (err) {
5269                                 ret = err;
5270                                 /*
5271                                  * We use &bio in above __extent_read_full_page,
5272                                  * so we ensure that if it returns error, the
5273                                  * current page fails to add itself to bio and
5274                                  * it's been unlocked.
5275                                  *
5276                                  * We must dec io_pages by ourselves.
5277                                  */
5278                                 atomic_dec(&eb->io_pages);
5279                         }
5280                 } else {
5281                         unlock_page(page);
5282                 }
5283         }
5284
5285         if (bio) {
5286                 err = submit_one_bio(bio, mirror_num, bio_flags);
5287                 if (err)
5288                         return err;
5289         }
5290
5291         if (ret || wait != WAIT_COMPLETE)
5292                 return ret;
5293
5294         for (i = 0; i < num_pages; i++) {
5295                 page = eb->pages[i];
5296                 wait_on_page_locked(page);
5297                 if (!PageUptodate(page))
5298                         ret = -EIO;
5299         }
5300
5301         return ret;
5302
5303 unlock_exit:
5304         while (locked_pages > 0) {
5305                 locked_pages--;
5306                 page = eb->pages[locked_pages];
5307                 unlock_page(page);
5308         }
5309         return ret;
5310 }
5311
5312 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5313                         unsigned long start,
5314                         unsigned long len)
5315 {
5316         size_t cur;
5317         size_t offset;
5318         struct page *page;
5319         char *kaddr;
5320         char *dst = (char *)dstv;
5321         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5322         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5323
5324         WARN_ON(start > eb->len);
5325         WARN_ON(start + len > eb->start + eb->len);
5326
5327         offset = (start_offset + start) & (PAGE_SIZE - 1);
5328
5329         while (len > 0) {
5330                 page = eb->pages[i];
5331
5332                 cur = min(len, (PAGE_SIZE - offset));
5333                 kaddr = page_address(page);
5334                 memcpy(dst, kaddr + offset, cur);
5335
5336                 dst += cur;
5337                 len -= cur;
5338                 offset = 0;
5339                 i++;
5340         }
5341 }
5342
5343 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5344                         unsigned long start,
5345                         unsigned long len)
5346 {
5347         size_t cur;
5348         size_t offset;
5349         struct page *page;
5350         char *kaddr;
5351         char __user *dst = (char __user *)dstv;
5352         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5353         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5354         int ret = 0;
5355
5356         WARN_ON(start > eb->len);
5357         WARN_ON(start + len > eb->start + eb->len);
5358
5359         offset = (start_offset + start) & (PAGE_SIZE - 1);
5360
5361         while (len > 0) {
5362                 page = eb->pages[i];
5363
5364                 cur = min(len, (PAGE_SIZE - offset));
5365                 kaddr = page_address(page);
5366                 if (copy_to_user(dst, kaddr + offset, cur)) {
5367                         ret = -EFAULT;
5368                         break;
5369                 }
5370
5371                 dst += cur;
5372                 len -= cur;
5373                 offset = 0;
5374                 i++;
5375         }
5376
5377         return ret;
5378 }
5379
5380 /*
5381  * return 0 if the item is found within a page.
5382  * return 1 if the item spans two pages.
5383  * return -EINVAL otherwise.
5384  */
5385 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5386                                unsigned long min_len, char **map,
5387                                unsigned long *map_start,
5388                                unsigned long *map_len)
5389 {
5390         size_t offset = start & (PAGE_SIZE - 1);
5391         char *kaddr;
5392         struct page *p;
5393         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5394         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5395         unsigned long end_i = (start_offset + start + min_len - 1) >>
5396                 PAGE_SHIFT;
5397
5398         if (i != end_i)
5399                 return 1;
5400
5401         if (i == 0) {
5402                 offset = start_offset;
5403                 *map_start = 0;
5404         } else {
5405                 offset = 0;
5406                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5407         }
5408
5409         if (start + min_len > eb->len) {
5410                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5411                        eb->start, eb->len, start, min_len);
5412                 return -EINVAL;
5413         }
5414
5415         p = eb->pages[i];
5416         kaddr = page_address(p);
5417         *map = kaddr + offset;
5418         *map_len = PAGE_SIZE - offset;
5419         return 0;
5420 }
5421
5422 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5423                           unsigned long start,
5424                           unsigned long len)
5425 {
5426         size_t cur;
5427         size_t offset;
5428         struct page *page;
5429         char *kaddr;
5430         char *ptr = (char *)ptrv;
5431         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5432         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5433         int ret = 0;
5434
5435         WARN_ON(start > eb->len);
5436         WARN_ON(start + len > eb->start + eb->len);
5437
5438         offset = (start_offset + start) & (PAGE_SIZE - 1);
5439
5440         while (len > 0) {
5441                 page = eb->pages[i];
5442
5443                 cur = min(len, (PAGE_SIZE - offset));
5444
5445                 kaddr = page_address(page);
5446                 ret = memcmp(ptr, kaddr + offset, cur);
5447                 if (ret)
5448                         break;
5449
5450                 ptr += cur;
5451                 len -= cur;
5452                 offset = 0;
5453                 i++;
5454         }
5455         return ret;
5456 }
5457
5458 void write_extent_buffer_chunk_tree_uuid(struct extent_buffer *eb,
5459                 const void *srcv)
5460 {
5461         char *kaddr;
5462
5463         WARN_ON(!PageUptodate(eb->pages[0]));
5464         kaddr = page_address(eb->pages[0]);
5465         memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv,
5466                         BTRFS_FSID_SIZE);
5467 }
5468
5469 void write_extent_buffer_fsid(struct extent_buffer *eb, const void *srcv)
5470 {
5471         char *kaddr;
5472
5473         WARN_ON(!PageUptodate(eb->pages[0]));
5474         kaddr = page_address(eb->pages[0]);
5475         memcpy(kaddr + offsetof(struct btrfs_header, fsid), srcv,
5476                         BTRFS_FSID_SIZE);
5477 }
5478
5479 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5480                          unsigned long start, unsigned long len)
5481 {
5482         size_t cur;
5483         size_t offset;
5484         struct page *page;
5485         char *kaddr;
5486         char *src = (char *)srcv;
5487         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5488         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5489
5490         WARN_ON(start > eb->len);
5491         WARN_ON(start + len > eb->start + eb->len);
5492
5493         offset = (start_offset + start) & (PAGE_SIZE - 1);
5494
5495         while (len > 0) {
5496                 page = eb->pages[i];
5497                 WARN_ON(!PageUptodate(page));
5498
5499                 cur = min(len, PAGE_SIZE - offset);
5500                 kaddr = page_address(page);
5501                 memcpy(kaddr + offset, src, cur);
5502
5503                 src += cur;
5504                 len -= cur;
5505                 offset = 0;
5506                 i++;
5507         }
5508 }
5509
5510 void memzero_extent_buffer(struct extent_buffer *eb, unsigned long start,
5511                 unsigned long len)
5512 {
5513         size_t cur;
5514         size_t offset;
5515         struct page *page;
5516         char *kaddr;
5517         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5518         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5519
5520         WARN_ON(start > eb->len);
5521         WARN_ON(start + len > eb->start + eb->len);
5522
5523         offset = (start_offset + start) & (PAGE_SIZE - 1);
5524
5525         while (len > 0) {
5526                 page = eb->pages[i];
5527                 WARN_ON(!PageUptodate(page));
5528
5529                 cur = min(len, PAGE_SIZE - offset);
5530                 kaddr = page_address(page);
5531                 memset(kaddr + offset, 0, cur);
5532
5533                 len -= cur;
5534                 offset = 0;
5535                 i++;
5536         }
5537 }
5538
5539 void copy_extent_buffer_full(struct extent_buffer *dst,
5540                              struct extent_buffer *src)
5541 {
5542         int i;
5543         unsigned num_pages;
5544
5545         ASSERT(dst->len == src->len);
5546
5547         num_pages = num_extent_pages(dst->start, dst->len);
5548         for (i = 0; i < num_pages; i++)
5549                 copy_page(page_address(dst->pages[i]),
5550                                 page_address(src->pages[i]));
5551 }
5552
5553 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5554                         unsigned long dst_offset, unsigned long src_offset,
5555                         unsigned long len)
5556 {
5557         u64 dst_len = dst->len;
5558         size_t cur;
5559         size_t offset;
5560         struct page *page;
5561         char *kaddr;
5562         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5563         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5564
5565         WARN_ON(src->len != dst_len);
5566
5567         offset = (start_offset + dst_offset) &
5568                 (PAGE_SIZE - 1);
5569
5570         while (len > 0) {
5571                 page = dst->pages[i];
5572                 WARN_ON(!PageUptodate(page));
5573
5574                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5575
5576                 kaddr = page_address(page);
5577                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5578
5579                 src_offset += cur;
5580                 len -= cur;
5581                 offset = 0;
5582                 i++;
5583         }
5584 }
5585
5586 void le_bitmap_set(u8 *map, unsigned int start, int len)
5587 {
5588         u8 *p = map + BIT_BYTE(start);
5589         const unsigned int size = start + len;
5590         int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5591         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);
5592
5593         while (len - bits_to_set >= 0) {
5594                 *p |= mask_to_set;
5595                 len -= bits_to_set;
5596                 bits_to_set = BITS_PER_BYTE;
5597                 mask_to_set = ~0;
5598                 p++;
5599         }
5600         if (len) {
5601                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5602                 *p |= mask_to_set;
5603         }
5604 }
5605
5606 void le_bitmap_clear(u8 *map, unsigned int start, int len)
5607 {
5608         u8 *p = map + BIT_BYTE(start);
5609         const unsigned int size = start + len;
5610         int bits_to_clear = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5611         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(start);
5612
5613         while (len - bits_to_clear >= 0) {
5614                 *p &= ~mask_to_clear;
5615                 len -= bits_to_clear;
5616                 bits_to_clear = BITS_PER_BYTE;
5617                 mask_to_clear = ~0;
5618                 p++;
5619         }
5620         if (len) {
5621                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5622                 *p &= ~mask_to_clear;
5623         }
5624 }
5625
5626 /*
5627  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5628  * given bit number
5629  * @eb: the extent buffer
5630  * @start: offset of the bitmap item in the extent buffer
5631  * @nr: bit number
5632  * @page_index: return index of the page in the extent buffer that contains the
5633  * given bit number
5634  * @page_offset: return offset into the page given by page_index
5635  *
5636  * This helper hides the ugliness of finding the byte in an extent buffer which
5637  * contains a given bit.
5638  */
5639 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5640                                     unsigned long start, unsigned long nr,
5641                                     unsigned long *page_index,
5642                                     size_t *page_offset)
5643 {
5644         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5645         size_t byte_offset = BIT_BYTE(nr);
5646         size_t offset;
5647
5648         /*
5649          * The byte we want is the offset of the extent buffer + the offset of
5650          * the bitmap item in the extent buffer + the offset of the byte in the
5651          * bitmap item.
5652          */
5653         offset = start_offset + start + byte_offset;
5654
5655         *page_index = offset >> PAGE_SHIFT;
5656         *page_offset = offset & (PAGE_SIZE - 1);
5657 }
5658
5659 /**
5660  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5661  * @eb: the extent buffer
5662  * @start: offset of the bitmap item in the extent buffer
5663  * @nr: bit number to test
5664  */
5665 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5666                            unsigned long nr)
5667 {
5668         u8 *kaddr;
5669         struct page *page;
5670         unsigned long i;
5671         size_t offset;
5672
5673         eb_bitmap_offset(eb, start, nr, &i, &offset);
5674         page = eb->pages[i];
5675         WARN_ON(!PageUptodate(page));
5676         kaddr = page_address(page);
5677         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5678 }
5679
5680 /**
5681  * extent_buffer_bitmap_set - set an area of a bitmap
5682  * @eb: the extent buffer
5683  * @start: offset of the bitmap item in the extent buffer
5684  * @pos: bit number of the first bit
5685  * @len: number of bits to set
5686  */
5687 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5688                               unsigned long pos, unsigned long len)
5689 {
5690         u8 *kaddr;
5691         struct page *page;
5692         unsigned long i;
5693         size_t offset;
5694         const unsigned int size = pos + len;
5695         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5696         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5697
5698         eb_bitmap_offset(eb, start, pos, &i, &offset);
5699         page = eb->pages[i];
5700         WARN_ON(!PageUptodate(page));
5701         kaddr = page_address(page);
5702
5703         while (len >= bits_to_set) {
5704                 kaddr[offset] |= mask_to_set;
5705                 len -= bits_to_set;
5706                 bits_to_set = BITS_PER_BYTE;
5707                 mask_to_set = ~0;
5708                 if (++offset >= PAGE_SIZE && len > 0) {
5709                         offset = 0;
5710                         page = eb->pages[++i];
5711                         WARN_ON(!PageUptodate(page));
5712                         kaddr = page_address(page);
5713                 }
5714         }
5715         if (len) {
5716                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5717                 kaddr[offset] |= mask_to_set;
5718         }
5719 }
5720
5721
5722 /**
5723  * extent_buffer_bitmap_clear - clear an area of a bitmap
5724  * @eb: the extent buffer
5725  * @start: offset of the bitmap item in the extent buffer
5726  * @pos: bit number of the first bit
5727  * @len: number of bits to clear
5728  */
5729 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5730                                 unsigned long pos, unsigned long len)
5731 {
5732         u8 *kaddr;
5733         struct page *page;
5734         unsigned long i;
5735         size_t offset;
5736         const unsigned int size = pos + len;
5737         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5738         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5739
5740         eb_bitmap_offset(eb, start, pos, &i, &offset);
5741         page = eb->pages[i];
5742         WARN_ON(!PageUptodate(page));
5743         kaddr = page_address(page);
5744
5745         while (len >= bits_to_clear) {
5746                 kaddr[offset] &= ~mask_to_clear;
5747                 len -= bits_to_clear;
5748                 bits_to_clear = BITS_PER_BYTE;
5749                 mask_to_clear = ~0;
5750                 if (++offset >= PAGE_SIZE && len > 0) {
5751                         offset = 0;
5752                         page = eb->pages[++i];
5753                         WARN_ON(!PageUptodate(page));
5754                         kaddr = page_address(page);
5755                 }
5756         }
5757         if (len) {
5758                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5759                 kaddr[offset] &= ~mask_to_clear;
5760         }
5761 }
5762
5763 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5764 {
5765         unsigned long distance = (src > dst) ? src - dst : dst - src;
5766         return distance < len;
5767 }
5768
5769 static void copy_pages(struct page *dst_page, struct page *src_page,
5770                        unsigned long dst_off, unsigned long src_off,
5771                        unsigned long len)
5772 {
5773         char *dst_kaddr = page_address(dst_page);
5774         char *src_kaddr;
5775         int must_memmove = 0;
5776
5777         if (dst_page != src_page) {
5778                 src_kaddr = page_address(src_page);
5779         } else {
5780                 src_kaddr = dst_kaddr;
5781                 if (areas_overlap(src_off, dst_off, len))
5782                         must_memmove = 1;
5783         }
5784
5785         if (must_memmove)
5786                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5787         else
5788                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5789 }
5790
5791 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5792                            unsigned long src_offset, unsigned long len)
5793 {
5794         struct btrfs_fs_info *fs_info = dst->fs_info;
5795         size_t cur;
5796         size_t dst_off_in_page;
5797         size_t src_off_in_page;
5798         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5799         unsigned long dst_i;
5800         unsigned long src_i;
5801
5802         if (src_offset + len > dst->len) {
5803                 btrfs_err(fs_info,
5804                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5805                          src_offset, len, dst->len);
5806                 BUG_ON(1);
5807         }
5808         if (dst_offset + len > dst->len) {
5809                 btrfs_err(fs_info,
5810                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5811                          dst_offset, len, dst->len);
5812                 BUG_ON(1);
5813         }
5814
5815         while (len > 0) {
5816                 dst_off_in_page = (start_offset + dst_offset) &
5817                         (PAGE_SIZE - 1);
5818                 src_off_in_page = (start_offset + src_offset) &
5819                         (PAGE_SIZE - 1);
5820
5821                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5822                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5823
5824                 cur = min(len, (unsigned long)(PAGE_SIZE -
5825                                                src_off_in_page));
5826                 cur = min_t(unsigned long, cur,
5827                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5828
5829                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5830                            dst_off_in_page, src_off_in_page, cur);
5831
5832                 src_offset += cur;
5833                 dst_offset += cur;
5834                 len -= cur;
5835         }
5836 }
5837
5838 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5839                            unsigned long src_offset, unsigned long len)
5840 {
5841         struct btrfs_fs_info *fs_info = dst->fs_info;
5842         size_t cur;
5843         size_t dst_off_in_page;
5844         size_t src_off_in_page;
5845         unsigned long dst_end = dst_offset + len - 1;
5846         unsigned long src_end = src_offset + len - 1;
5847         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5848         unsigned long dst_i;
5849         unsigned long src_i;
5850
5851         if (src_offset + len > dst->len) {
5852                 btrfs_err(fs_info,
5853                           "memmove bogus src_offset %lu move len %lu len %lu",
5854                           src_offset, len, dst->len);
5855                 BUG_ON(1);
5856         }
5857         if (dst_offset + len > dst->len) {
5858                 btrfs_err(fs_info,
5859                           "memmove bogus dst_offset %lu move len %lu len %lu",
5860                           dst_offset, len, dst->len);
5861                 BUG_ON(1);
5862         }
5863         if (dst_offset < src_offset) {
5864                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5865                 return;
5866         }
5867         while (len > 0) {
5868                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5869                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5870
5871                 dst_off_in_page = (start_offset + dst_end) &
5872                         (PAGE_SIZE - 1);
5873                 src_off_in_page = (start_offset + src_end) &
5874                         (PAGE_SIZE - 1);
5875
5876                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5877                 cur = min(cur, dst_off_in_page + 1);
5878                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5879                            dst_off_in_page - cur + 1,
5880                            src_off_in_page - cur + 1, cur);
5881
5882                 dst_end -= cur;
5883                 src_end -= cur;
5884                 len -= cur;
5885         }
5886 }
5887
5888 int try_release_extent_buffer(struct page *page)
5889 {
5890         struct extent_buffer *eb;
5891
5892         /*
5893          * We need to make sure nobody is attaching this page to an eb right
5894          * now.
5895          */
5896         spin_lock(&page->mapping->private_lock);
5897         if (!PagePrivate(page)) {
5898                 spin_unlock(&page->mapping->private_lock);
5899                 return 1;
5900         }
5901
5902         eb = (struct extent_buffer *)page->private;
5903         BUG_ON(!eb);
5904
5905         /*
5906          * This is a little awful but should be ok, we need to make sure that
5907          * the eb doesn't disappear out from under us while we're looking at
5908          * this page.
5909          */
5910         spin_lock(&eb->refs_lock);
5911         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5912                 spin_unlock(&eb->refs_lock);
5913                 spin_unlock(&page->mapping->private_lock);
5914                 return 0;
5915         }
5916         spin_unlock(&page->mapping->private_lock);
5917
5918         /*
5919          * If tree ref isn't set then we know the ref on this eb is a real ref,
5920          * so just return, this page will likely be freed soon anyway.
5921          */
5922         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5923                 spin_unlock(&eb->refs_lock);
5924                 return 0;
5925         }
5926
5927         return release_extent_buffer(eb);
5928 }