fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/spinlock.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/swap.h>
  10 #include <linux/writeback.h>
  11 #include <linux/pagevec.h>
  12 #include <linux/prefetch.h>
  13 #include <linux/cleancache.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "ctree.h"
  17 #include "btrfs_inode.h"
  18 #include "volumes.h"
  19 #include "check-integrity.h"
  20 #include "locking.h"
  21 #include "rcu-string.h"
  22 #include "backref.h"
  23
  24 static struct kmem_cache *extent_state_cache;
  25 static struct kmem_cache *extent_buffer_cache;
  26 static struct bio_set *btrfs_bioset;
  27
  28 static inline bool extent_state_in_tree(const struct extent_state *state)
  29 {
  30         return !RB_EMPTY_NODE(&state->rb_node);
  31 }
  32
  33 #ifdef CONFIG_BTRFS_DEBUG
  34 static LIST_HEAD(buffers);
  35 static LIST_HEAD(states);
  36
  37 static DEFINE_SPINLOCK(leak_lock);
  38
  39 static inline
  40 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  41 {
  42         unsigned long flags;
  43
  44         spin_lock_irqsave(&leak_lock, flags);
  45         list_add(new, head);
  46         spin_unlock_irqrestore(&leak_lock, flags);
  47 }
  48
  49 static inline
  50 void btrfs_leak_debug_del(struct list_head *entry)
  51 {
  52         unsigned long flags;
  53
  54         spin_lock_irqsave(&leak_lock, flags);
  55         list_del(entry);
  56         spin_unlock_irqrestore(&leak_lock, flags);
  57 }
  58
  59 static inline
  60 void btrfs_leak_debug_check(void)
  61 {
  62         struct extent_state *state;
  63         struct extent_buffer *eb;
  64
  65         while (!list_empty(&states)) {
  66                 state = list_entry(states.next, struct extent_state, leak_list);
  67                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  68                        state->start, state->end, state->state,
  69                        extent_state_in_tree(state),
  70                        atomic_read(&state->refs));
  71                 list_del(&state->leak_list);
  72                 kmem_cache_free(extent_state_cache, state);
  73         }
  74
  75         while (!list_empty(&buffers)) {
  76                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  77                 printk(KERN_ERR "BTRFS: buffer leak start %llu len %lu "
  78                        "refs %d\n",
  79                        eb->start, eb->len, atomic_read(&eb->refs));
  80                 list_del(&eb->leak_list);
  81                 kmem_cache_free(extent_buffer_cache, eb);
  82         }
  83 }
  84
  85 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  86         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  87 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  88                 struct extent_io_tree *tree, u64 start, u64 end)
  89 {
  90         struct inode *inode;
  91         u64 isize;
  92
  93         if (!tree->mapping)
  94                 return;
  95
  96         inode = tree->mapping->host;
  97         isize = i_size_read(inode);
  98         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
  99                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
 100                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
 101                                 caller, btrfs_ino(inode), isize, start, end);
 102         }
 103 }
 104 #else
 105 #define btrfs_leak_debug_add(new, head) do {} while (0)
 106 #define btrfs_leak_debug_del(entry)     do {} while (0)
 107 #define btrfs_leak_debug_check()        do {} while (0)
 108 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 109 #endif
 110
 111 #define BUFFER_LRU_MAX 64
 112
 113 struct tree_entry {
 114         u64 start;
 115         u64 end;
 116         struct rb_node rb_node;
 117 };
 118
 119 struct extent_page_data {
 120         struct bio *bio;
 121         struct extent_io_tree *tree;
 122         get_extent_t *get_extent;
 123         unsigned long bio_flags;
 124
 125         /* tells writepage not to lock the state bits for this range
 126          * it still does the unlocking
 127          */
 128         unsigned int extent_locked:1;
 129
 130         /* tells the submit_bio code to use a WRITE_SYNC */
 131         unsigned int sync_io:1;
 132 };
 133
 134 static void add_extent_changeset(struct extent_state *state, unsigned bits,
 135                                  struct extent_changeset *changeset,
 136                                  int set)
 137 {
 138         int ret;
 139
 140         if (!changeset)
 141                 return;
 142         if (set && (state->state & bits) == bits)
 143                 return;
 144         if (!set && (state->state & bits) == 0)
 145                 return;
 146         changeset->bytes_changed += state->end - state->start + 1;
 147         ret = ulist_add(changeset->range_changed, state->start, state->end,
 148                         GFP_ATOMIC);
 149         /* ENOMEM */
 150         BUG_ON(ret < 0);
 151 }
 152
 153 static noinline void flush_write_bio(void *data);
 154 static inline struct btrfs_fs_info *
 155 tree_fs_info(struct extent_io_tree *tree)
 156 {
 157         if (!tree->mapping)
 158                 return NULL;
 159         return btrfs_sb(tree->mapping->host->i_sb);
 160 }
 161
 162 int __init extent_io_init(void)
 163 {
 164         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 165                         sizeof(struct extent_state), 0,
 166                         SLAB_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         state = kmem_cache_alloc(extent_state_cache, mask);
 230         if (!state)
 231                 return state;
 232         state->state = 0;
 233         state->failrec = NULL;
 234         RB_CLEAR_NODE(&state->rb_node);
 235         btrfs_leak_debug_add(&state->leak_list, &states);
 236         atomic_set(&state->refs, 1);
 237         init_waitqueue_head(&state->wq);
 238         trace_alloc_extent_state(state, mask, _RET_IP_);
 239         return state;
 240 }
 241
 242 void free_extent_state(struct extent_state *state)
 243 {
 244         if (!state)
 245                 return;
 246         if (atomic_dec_and_test(&state->refs)) {
 247                 WARN_ON(extent_state_in_tree(state));
 248                 btrfs_leak_debug_del(&state->leak_list);
 249                 trace_free_extent_state(state, _RET_IP_);
 250                 kmem_cache_free(extent_state_cache, state);
 251         }
 252 }
 253
 254 static struct rb_node *tree_insert(struct rb_root *root,
 255                                    struct rb_node *search_start,
 256                                    u64 offset,
 257                                    struct rb_node *node,
 258                                    struct rb_node ***p_in,
 259                                    struct rb_node **parent_in)
 260 {
 261         struct rb_node **p;
 262         struct rb_node *parent = NULL;
 263         struct tree_entry *entry;
 264
 265         if (p_in && parent_in) {
 266                 p = *p_in;
 267                 parent = *parent_in;
 268                 goto do_insert;
 269         }
 270
 271         p = search_start ? &search_start : &root->rb_node;
 272         while (*p) {
 273                 parent = *p;
 274                 entry = rb_entry(parent, struct tree_entry, rb_node);
 275
 276                 if (offset < entry->start)
 277                         p = &(*p)->rb_left;
 278                 else if (offset > entry->end)
 279                         p = &(*p)->rb_right;
 280                 else
 281                         return parent;
 282         }
 283
 284 do_insert:
 285         rb_link_node(node, parent, p);
 286         rb_insert_color(node, root);
 287         return NULL;
 288 }
 289
 290 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 291                                       struct rb_node **prev_ret,
 292                                       struct rb_node **next_ret,
 293                                       struct rb_node ***p_ret,
 294                                       struct rb_node **parent_ret)
 295 {
 296         struct rb_root *root = &tree->state;
 297         struct rb_node **n = &root->rb_node;
 298         struct rb_node *prev = NULL;
 299         struct rb_node *orig_prev = NULL;
 300         struct tree_entry *entry;
 301         struct tree_entry *prev_entry = NULL;
 302
 303         while (*n) {
 304                 prev = *n;
 305                 entry = rb_entry(prev, struct tree_entry, rb_node);
 306                 prev_entry = entry;
 307
 308                 if (offset < entry->start)
 309                         n = &(*n)->rb_left;
 310                 else if (offset > entry->end)
 311                         n = &(*n)->rb_right;
 312                 else
 313                         return *n;
 314         }
 315
 316         if (p_ret)
 317                 *p_ret = n;
 318         if (parent_ret)
 319                 *parent_ret = prev;
 320
 321         if (prev_ret) {
 322                 orig_prev = prev;
 323                 while (prev && offset > prev_entry->end) {
 324                         prev = rb_next(prev);
 325                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 326                 }
 327                 *prev_ret = prev;
 328                 prev = orig_prev;
 329         }
 330
 331         if (next_ret) {
 332                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 333                 while (prev && offset < prev_entry->start) {
 334                         prev = rb_prev(prev);
 335                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 336                 }
 337                 *next_ret = prev;
 338         }
 339         return NULL;
 340 }
 341
 342 static inline struct rb_node *
 343 tree_search_for_insert(struct extent_io_tree *tree,
 344                        u64 offset,
 345                        struct rb_node ***p_ret,
 346                        struct rb_node **parent_ret)
 347 {
 348         struct rb_node *prev = NULL;
 349         struct rb_node *ret;
 350
 351         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 352         if (!ret)
 353                 return prev;
 354         return ret;
 355 }
 356
 357 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 358                                           u64 offset)
 359 {
 360         return tree_search_for_insert(tree, offset, NULL, NULL);
 361 }
 362
 363 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 364                      struct extent_state *other)
 365 {
 366         if (tree->ops && tree->ops->merge_extent_hook)
 367                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 368                                              other);
 369 }
 370
 371 /*
 372  * utility function to look for merge candidates inside a given range.
 373  * Any extents with matching state are merged together into a single
 374  * extent in the tree.  Extents with EXTENT_IO in their state field
 375  * are not merged because the end_io handlers need to be able to do
 376  * operations on them without sleeping (or doing allocations/splits).
 377  *
 378  * This should be called with the tree lock held.
 379  */
 380 static void merge_state(struct extent_io_tree *tree,
 381                         struct extent_state *state)
 382 {
 383         struct extent_state *other;
 384         struct rb_node *other_node;
 385
 386         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 387                 return;
 388
 389         other_node = rb_prev(&state->rb_node);
 390         if (other_node) {
 391                 other = rb_entry(other_node, struct extent_state, rb_node);
 392                 if (other->end == state->start - 1 &&
 393                     other->state == state->state) {
 394                         merge_cb(tree, state, other);
 395                         state->start = other->start;
 396                         rb_erase(&other->rb_node, &tree->state);
 397                         RB_CLEAR_NODE(&other->rb_node);
 398                         free_extent_state(other);
 399                 }
 400         }
 401         other_node = rb_next(&state->rb_node);
 402         if (other_node) {
 403                 other = rb_entry(other_node, struct extent_state, rb_node);
 404                 if (other->start == state->end + 1 &&
 405                     other->state == state->state) {
 406                         merge_cb(tree, state, other);
 407                         state->end = other->end;
 408                         rb_erase(&other->rb_node, &tree->state);
 409                         RB_CLEAR_NODE(&other->rb_node);
 410                         free_extent_state(other);
 411                 }
 412         }
 413 }
 414
 415 static void set_state_cb(struct extent_io_tree *tree,
 416                          struct extent_state *state, unsigned *bits)
 417 {
 418         if (tree->ops && tree->ops->set_bit_hook)
 419                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 420 }
 421
 422 static void clear_state_cb(struct extent_io_tree *tree,
 423                            struct extent_state *state, unsigned *bits)
 424 {
 425         if (tree->ops && tree->ops->clear_bit_hook)
 426                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 427 }
 428
 429 static void set_state_bits(struct extent_io_tree *tree,
 430                            struct extent_state *state, unsigned *bits,
 431                            struct extent_changeset *changeset);
 432
 433 /*
 434  * insert an extent_state struct into the tree.  'bits' are set on the
 435  * struct before it is inserted.
 436  *
 437  * This may return -EEXIST if the extent is already there, in which case the
 438  * state struct is freed.
 439  *
 440  * The tree lock is not taken internally.  This is a utility function and
 441  * probably isn't what you want to call (see set/clear_extent_bit).
 442  */
 443 static int insert_state(struct extent_io_tree *tree,
 444                         struct extent_state *state, u64 start, u64 end,
 445                         struct rb_node ***p,
 446                         struct rb_node **parent,
 447                         unsigned *bits, struct extent_changeset *changeset)
 448 {
 449         struct rb_node *node;
 450
 451         if (end < start)
 452                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 453                        end, start);
 454         state->start = start;
 455         state->end = end;
 456
 457         set_state_bits(tree, state, bits, changeset);
 458
 459         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 460         if (node) {
 461                 struct extent_state *found;
 462                 found = rb_entry(node, struct extent_state, rb_node);
 463                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
 464                        "%llu %llu\n",
 465                        found->start, found->end, start, end);
 466                 return -EEXIST;
 467         }
 468         merge_state(tree, state);
 469         return 0;
 470 }
 471
 472 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 473                      u64 split)
 474 {
 475         if (tree->ops && tree->ops->split_extent_hook)
 476                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 477 }
 478
 479 /*
 480  * split a given extent state struct in two, inserting the preallocated
 481  * struct 'prealloc' as the newly created second half.  'split' indicates an
 482  * offset inside 'orig' where it should be split.
 483  *
 484  * Before calling,
 485  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 486  * are two extent state structs in the tree:
 487  * prealloc: [orig->start, split - 1]
 488  * orig: [ split, orig->end ]
 489  *
 490  * The tree locks are not taken by this function. They need to be held
 491  * by the caller.
 492  */
 493 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 494                        struct extent_state *prealloc, u64 split)
 495 {
 496         struct rb_node *node;
 497
 498         split_cb(tree, orig, split);
 499
 500         prealloc->start = orig->start;
 501         prealloc->end = split - 1;
 502         prealloc->state = orig->state;
 503         orig->start = split;
 504
 505         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 506                            &prealloc->rb_node, NULL, NULL);
 507         if (node) {
 508                 free_extent_state(prealloc);
 509                 return -EEXIST;
 510         }
 511         return 0;
 512 }
 513
 514 static struct extent_state *next_state(struct extent_state *state)
 515 {
 516         struct rb_node *next = rb_next(&state->rb_node);
 517         if (next)
 518                 return rb_entry(next, struct extent_state, rb_node);
 519         else
 520                 return NULL;
 521 }
 522
 523 /*
 524  * utility function to clear some bits in an extent state struct.
 525  * it will optionally wake up any one waiting on this state (wake == 1).
 526  *
 527  * If no bits are set on the state struct after clearing things, the
 528  * struct is freed and removed from the tree
 529  */
 530 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 531                                             struct extent_state *state,
 532                                             unsigned *bits, int wake,
 533                                             struct extent_changeset *changeset)
 534 {
 535         struct extent_state *next;
 536         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 537
 538         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 539                 u64 range = state->end - state->start + 1;
 540                 WARN_ON(range > tree->dirty_bytes);
 541                 tree->dirty_bytes -= range;
 542         }
 543         clear_state_cb(tree, state, bits);
 544         add_extent_changeset(state, bits_to_clear, changeset, 0);
 545         state->state &= ~bits_to_clear;
 546         if (wake)
 547                 wake_up(&state->wq);
 548         if (state->state == 0) {
 549                 next = next_state(state);
 550                 if (extent_state_in_tree(state)) {
 551                         rb_erase(&state->rb_node, &tree->state);
 552                         RB_CLEAR_NODE(&state->rb_node);
 553                         free_extent_state(state);
 554                 } else {
 555                         WARN_ON(1);
 556                 }
 557         } else {
 558                 merge_state(tree, state);
 559                 next = next_state(state);
 560         }
 561         return next;
 562 }
 563
 564 static struct extent_state *
 565 alloc_extent_state_atomic(struct extent_state *prealloc)
 566 {
 567         if (!prealloc)
 568                 prealloc = alloc_extent_state(GFP_ATOMIC);
 569
 570         return prealloc;
 571 }
 572
 573 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 574 {
 575         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 576                     "Extent tree was modified by another "
 577                     "thread while locked.");
 578 }
 579
 580 /*
 581  * clear some bits on a range in the tree.  This may require splitting
 582  * or inserting elements in the tree, so the gfp mask is used to
 583  * indicate which allocations or sleeping are allowed.
 584  *
 585  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 586  * the given range from the tree regardless of state (ie for truncate).
 587  *
 588  * the range [start, end] is inclusive.
 589  *
 590  * This takes the tree lock, and returns 0 on success and < 0 on error.
 591  */
 592 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 593                               unsigned bits, int wake, int delete,
 594                               struct extent_state **cached_state,
 595                               gfp_t mask, struct extent_changeset *changeset)
 596 {
 597         struct extent_state *state;
 598         struct extent_state *cached;
 599         struct extent_state *prealloc = NULL;
 600         struct rb_node *node;
 601         u64 last_end;
 602         int err;
 603         int clear = 0;
 604
 605         btrfs_debug_check_extent_io_range(tree, start, end);
 606
 607         if (bits & EXTENT_DELALLOC)
 608                 bits |= EXTENT_NORESERVE;
 609
 610         if (delete)
 611                 bits |= ~EXTENT_CTLBITS;
 612         bits |= EXTENT_FIRST_DELALLOC;
 613
 614         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 615                 clear = 1;
 616 again:
 617         if (!prealloc && gfpflags_allow_blocking(mask)) {
 618                 /*
 619                  * Don't care for allocation failure here because we might end
 620                  * up not needing the pre-allocated extent state at all, which
 621                  * is the case if we only have in the tree extent states that
 622                  * cover our input range and don't cover too any other range.
 623                  * If we end up needing a new extent state we allocate it later.
 624                  */
 625                 prealloc = alloc_extent_state(mask);
 626         }
 627
 628         spin_lock(&tree->lock);
 629         if (cached_state) {
 630                 cached = *cached_state;
 631
 632                 if (clear) {
 633                         *cached_state = NULL;
 634                         cached_state = NULL;
 635                 }
 636
 637                 if (cached && extent_state_in_tree(cached) &&
 638                     cached->start <= start && cached->end > start) {
 639                         if (clear)
 640                                 atomic_dec(&cached->refs);
 641                         state = cached;
 642                         goto hit_next;
 643                 }
 644                 if (clear)
 645                         free_extent_state(cached);
 646         }
 647         /*
 648          * this search will find the extents that end after
 649          * our range starts
 650          */
 651         node = tree_search(tree, start);
 652         if (!node)
 653                 goto out;
 654         state = rb_entry(node, struct extent_state, rb_node);
 655 hit_next:
 656         if (state->start > end)
 657                 goto out;
 658         WARN_ON(state->end < start);
 659         last_end = state->end;
 660
 661         /* the state doesn't have the wanted bits, go ahead */
 662         if (!(state->state & bits)) {
 663                 state = next_state(state);
 664                 goto next;
 665         }
 666
 667         /*
 668          *     | ---- desired range ---- |
 669          *  | state | or
 670          *  | ------------- state -------------- |
 671          *
 672          * We need to split the extent we found, and may flip
 673          * bits on second half.
 674          *
 675          * If the extent we found extends past our range, we
 676          * just split and search again.  It'll get split again
 677          * the next time though.
 678          *
 679          * If the extent we found is inside our range, we clear
 680          * the desired bit on it.
 681          */
 682
 683         if (state->start < start) {
 684                 prealloc = alloc_extent_state_atomic(prealloc);
 685                 BUG_ON(!prealloc);
 686                 err = split_state(tree, state, prealloc, start);
 687                 if (err)
 688                         extent_io_tree_panic(tree, err);
 689
 690                 prealloc = NULL;
 691                 if (err)
 692                         goto out;
 693                 if (state->end <= end) {
 694                         state = clear_state_bit(tree, state, &bits, wake,
 695                                                 changeset);
 696                         goto next;
 697                 }
 698                 goto search_again;
 699         }
 700         /*
 701          * | ---- desired range ---- |
 702          *                        | state |
 703          * We need to split the extent, and clear the bit
 704          * on the first half
 705          */
 706         if (state->start <= end && state->end > end) {
 707                 prealloc = alloc_extent_state_atomic(prealloc);
 708                 BUG_ON(!prealloc);
 709                 err = split_state(tree, state, prealloc, end + 1);
 710                 if (err)
 711                         extent_io_tree_panic(tree, err);
 712
 713                 if (wake)
 714                         wake_up(&state->wq);
 715
 716                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 717
 718                 prealloc = NULL;
 719                 goto out;
 720         }
 721
 722         state = clear_state_bit(tree, state, &bits, wake, changeset);
 723 next:
 724         if (last_end == (u64)-1)
 725                 goto out;
 726         start = last_end + 1;
 727         if (start <= end && state && !need_resched())
 728                 goto hit_next;
 729
 730 search_again:
 731         if (start > end)
 732                 goto out;
 733         spin_unlock(&tree->lock);
 734         if (gfpflags_allow_blocking(mask))
 735                 cond_resched();
 736         goto again;
 737
 738 out:
 739         spin_unlock(&tree->lock);
 740         if (prealloc)
 741                 free_extent_state(prealloc);
 742
 743         return 0;
 744
 745 }
 746
 747 static void wait_on_state(struct extent_io_tree *tree,
 748                           struct extent_state *state)
 749                 __releases(tree->lock)
 750                 __acquires(tree->lock)
 751 {
 752         DEFINE_WAIT(wait);
 753         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 754         spin_unlock(&tree->lock);
 755         schedule();
 756         spin_lock(&tree->lock);
 757         finish_wait(&state->wq, &wait);
 758 }
 759
 760 /*
 761  * waits for one or more bits to clear on a range in the state tree.
 762  * The range [start, end] is inclusive.
 763  * The tree lock is taken by this function
 764  */
 765 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 766                             unsigned long bits)
 767 {
 768         struct extent_state *state;
 769         struct rb_node *node;
 770
 771         btrfs_debug_check_extent_io_range(tree, start, end);
 772
 773         spin_lock(&tree->lock);
 774 again:
 775         while (1) {
 776                 /*
 777                  * this search will find all the extents that end after
 778                  * our range starts
 779                  */
 780                 node = tree_search(tree, start);
 781 process_node:
 782                 if (!node)
 783                         break;
 784
 785                 state = rb_entry(node, struct extent_state, rb_node);
 786
 787                 if (state->start > end)
 788                         goto out;
 789
 790                 if (state->state & bits) {
 791                         start = state->start;
 792                         atomic_inc(&state->refs);
 793                         wait_on_state(tree, state);
 794                         free_extent_state(state);
 795                         goto again;
 796                 }
 797                 start = state->end + 1;
 798
 799                 if (start > end)
 800                         break;
 801
 802                 if (!cond_resched_lock(&tree->lock)) {
 803                         node = rb_next(node);
 804                         goto process_node;
 805                 }
 806         }
 807 out:
 808         spin_unlock(&tree->lock);
 809 }
 810
 811 static void set_state_bits(struct extent_io_tree *tree,
 812                            struct extent_state *state,
 813                            unsigned *bits, struct extent_changeset *changeset)
 814 {
 815         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 816
 817         set_state_cb(tree, state, bits);
 818         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 819                 u64 range = state->end - state->start + 1;
 820                 tree->dirty_bytes += range;
 821         }
 822         add_extent_changeset(state, bits_to_set, changeset, 1);
 823         state->state |= bits_to_set;
 824 }
 825
 826 static void cache_state_if_flags(struct extent_state *state,
 827                                  struct extent_state **cached_ptr,
 828                                  unsigned flags)
 829 {
 830         if (cached_ptr && !(*cached_ptr)) {
 831                 if (!flags || (state->state & flags)) {
 832                         *cached_ptr = state;
 833                         atomic_inc(&state->refs);
 834                 }
 835         }
 836 }
 837
 838 static void cache_state(struct extent_state *state,
 839                         struct extent_state **cached_ptr)
 840 {
 841         return cache_state_if_flags(state, cached_ptr,
 842                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 843 }
 844
 845 /*
 846  * set some bits on a range in the tree.  This may require allocations or
 847  * sleeping, so the gfp mask is used to indicate what is allowed.
 848  *
 849  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 850  * part of the range already has the desired bits set.  The start of the
 851  * existing range is returned in failed_start in this case.
 852  *
 853  * [start, end] is inclusive This takes the tree lock.
 854  */
 855
 856 static int __must_check
 857 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 858                  unsigned bits, unsigned exclusive_bits,
 859                  u64 *failed_start, struct extent_state **cached_state,
 860                  gfp_t mask, struct extent_changeset *changeset)
 861 {
 862         struct extent_state *state;
 863         struct extent_state *prealloc = NULL;
 864         struct rb_node *node;
 865         struct rb_node **p;
 866         struct rb_node *parent;
 867         int err = 0;
 868         u64 last_start;
 869         u64 last_end;
 870
 871         btrfs_debug_check_extent_io_range(tree, start, end);
 872
 873         bits |= EXTENT_FIRST_DELALLOC;
 874 again:
 875         if (!prealloc && gfpflags_allow_blocking(mask)) {
 876                 /*
 877                  * Don't care for allocation failure here because we might end
 878                  * up not needing the pre-allocated extent state at all, which
 879                  * is the case if we only have in the tree extent states that
 880                  * cover our input range and don't cover too any other range.
 881                  * If we end up needing a new extent state we allocate it later.
 882                  */
 883                 prealloc = alloc_extent_state(mask);
 884         }
 885
 886         spin_lock(&tree->lock);
 887         if (cached_state && *cached_state) {
 888                 state = *cached_state;
 889                 if (state->start <= start && state->end > start &&
 890                     extent_state_in_tree(state)) {
 891                         node = &state->rb_node;
 892                         goto hit_next;
 893                 }
 894         }
 895         /*
 896          * this search will find all the extents that end after
 897          * our range starts.
 898          */
 899         node = tree_search_for_insert(tree, start, &p, &parent);
 900         if (!node) {
 901                 prealloc = alloc_extent_state_atomic(prealloc);
 902                 BUG_ON(!prealloc);
 903                 err = insert_state(tree, prealloc, start, end,
 904                                    &p, &parent, &bits, changeset);
 905                 if (err)
 906                         extent_io_tree_panic(tree, err);
 907
 908                 cache_state(prealloc, cached_state);
 909                 prealloc = NULL;
 910                 goto out;
 911         }
 912         state = rb_entry(node, struct extent_state, rb_node);
 913 hit_next:
 914         last_start = state->start;
 915         last_end = state->end;
 916
 917         /*
 918          * | ---- desired range ---- |
 919          * | state |
 920          *
 921          * Just lock what we found and keep going
 922          */
 923         if (state->start == start && state->end <= end) {
 924                 if (state->state & exclusive_bits) {
 925                         *failed_start = state->start;
 926                         err = -EEXIST;
 927                         goto out;
 928                 }
 929
 930                 set_state_bits(tree, state, &bits, changeset);
 931                 cache_state(state, cached_state);
 932                 merge_state(tree, state);
 933                 if (last_end == (u64)-1)
 934                         goto out;
 935                 start = last_end + 1;
 936                 state = next_state(state);
 937                 if (start < end && state && state->start == start &&
 938                     !need_resched())
 939                         goto hit_next;
 940                 goto search_again;
 941         }
 942
 943         /*
 944          *     | ---- desired range ---- |
 945          * | state |
 946          *   or
 947          * | ------------- state -------------- |
 948          *
 949          * We need to split the extent we found, and may flip bits on
 950          * second half.
 951          *
 952          * If the extent we found extends past our
 953          * range, we just split and search again.  It'll get split
 954          * again the next time though.
 955          *
 956          * If the extent we found is inside our range, we set the
 957          * desired bit on it.
 958          */
 959         if (state->start < start) {
 960                 if (state->state & exclusive_bits) {
 961                         *failed_start = start;
 962                         err = -EEXIST;
 963                         goto out;
 964                 }
 965
 966                 prealloc = alloc_extent_state_atomic(prealloc);
 967                 BUG_ON(!prealloc);
 968                 err = split_state(tree, state, prealloc, start);
 969                 if (err)
 970                         extent_io_tree_panic(tree, err);
 971
 972                 prealloc = NULL;
 973                 if (err)
 974                         goto out;
 975                 if (state->end <= end) {
 976                         set_state_bits(tree, state, &bits, changeset);
 977                         cache_state(state, cached_state);
 978                         merge_state(tree, state);
 979                         if (last_end == (u64)-1)
 980                                 goto out;
 981                         start = last_end + 1;
 982                         state = next_state(state);
 983                         if (start < end && state && state->start == start &&
 984                             !need_resched())
 985                                 goto hit_next;
 986                 }
 987                 goto search_again;
 988         }
 989         /*
 990          * | ---- desired range ---- |
 991          *     | state | or               | state |
 992          *
 993          * There's a hole, we need to insert something in it and
 994          * ignore the extent we found.
 995          */
 996         if (state->start > start) {
 997                 u64 this_end;
 998                 if (end < last_start)
 999                         this_end = end;
1000                 else
1001                         this_end = last_start - 1;
1002
1003                 prealloc = alloc_extent_state_atomic(prealloc);
1004                 BUG_ON(!prealloc);
1005
1006                 /*
1007                  * Avoid to free 'prealloc' if it can be merged with
1008                  * the later extent.
1009                  */
1010                 err = insert_state(tree, prealloc, start, this_end,
1011                                    NULL, NULL, &bits, changeset);
1012                 if (err)
1013                         extent_io_tree_panic(tree, err);
1014
1015                 cache_state(prealloc, cached_state);
1016                 prealloc = NULL;
1017                 start = this_end + 1;
1018                 goto search_again;
1019         }
1020         /*
1021          * | ---- desired range ---- |
1022          *                        | state |
1023          * We need to split the extent, and set the bit
1024          * on the first half
1025          */
1026         if (state->start <= end && state->end > end) {
1027                 if (state->state & exclusive_bits) {
1028                         *failed_start = start;
1029                         err = -EEXIST;
1030                         goto out;
1031                 }
1032
1033                 prealloc = alloc_extent_state_atomic(prealloc);
1034                 BUG_ON(!prealloc);
1035                 err = split_state(tree, state, prealloc, end + 1);
1036                 if (err)
1037                         extent_io_tree_panic(tree, err);
1038
1039                 set_state_bits(tree, prealloc, &bits, changeset);
1040                 cache_state(prealloc, cached_state);
1041                 merge_state(tree, prealloc);
1042                 prealloc = NULL;
1043                 goto out;
1044         }
1045
1046 search_again:
1047         if (start > end)
1048                 goto out;
1049         spin_unlock(&tree->lock);
1050         if (gfpflags_allow_blocking(mask))
1051                 cond_resched();
1052         goto again;
1053
1054 out:
1055         spin_unlock(&tree->lock);
1056         if (prealloc)
1057                 free_extent_state(prealloc);
1058
1059         return err;
1060
1061 }
1062
1063 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1064                    unsigned bits, u64 * failed_start,
1065                    struct extent_state **cached_state, gfp_t mask)
1066 {
1067         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1068                                 cached_state, mask, NULL);
1069 }
1070
1071
1072 /**
1073  * convert_extent_bit - convert all bits in a given range from one bit to
1074  *                      another
1075  * @tree:       the io tree to search
1076  * @start:      the start offset in bytes
1077  * @end:        the end offset in bytes (inclusive)
1078  * @bits:       the bits to set in this range
1079  * @clear_bits: the bits to clear in this range
1080  * @cached_state:       state that we're going to cache
1081  *
1082  * This will go through and set bits for the given range.  If any states exist
1083  * already in this range they are set with the given bit and cleared of the
1084  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1085  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1086  * boundary bits like LOCK.
1087  *
1088  * All allocations are done with GFP_NOFS.
1089  */
1090 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1091                        unsigned bits, unsigned clear_bits,
1092                        struct extent_state **cached_state)
1093 {
1094         struct extent_state *state;
1095         struct extent_state *prealloc = NULL;
1096         struct rb_node *node;
1097         struct rb_node **p;
1098         struct rb_node *parent;
1099         int err = 0;
1100         u64 last_start;
1101         u64 last_end;
1102         bool first_iteration = true;
1103
1104         btrfs_debug_check_extent_io_range(tree, start, end);
1105
1106 again:
1107         if (!prealloc) {
1108                 /*
1109                  * Best effort, don't worry if extent state allocation fails
1110                  * here for the first iteration. We might have a cached state
1111                  * that matches exactly the target range, in which case no
1112                  * extent state allocations are needed. We'll only know this
1113                  * after locking the tree.
1114                  */
1115                 prealloc = alloc_extent_state(GFP_NOFS);
1116                 if (!prealloc && !first_iteration)
1117                         return -ENOMEM;
1118         }
1119
1120         spin_lock(&tree->lock);
1121         if (cached_state && *cached_state) {
1122                 state = *cached_state;
1123                 if (state->start <= start && state->end > start &&
1124                     extent_state_in_tree(state)) {
1125                         node = &state->rb_node;
1126                         goto hit_next;
1127                 }
1128         }
1129
1130         /*
1131          * this search will find all the extents that end after
1132          * our range starts.
1133          */
1134         node = tree_search_for_insert(tree, start, &p, &parent);
1135         if (!node) {
1136                 prealloc = alloc_extent_state_atomic(prealloc);
1137                 if (!prealloc) {
1138                         err = -ENOMEM;
1139                         goto out;
1140                 }
1141                 err = insert_state(tree, prealloc, start, end,
1142                                    &p, &parent, &bits, NULL);
1143                 if (err)
1144                         extent_io_tree_panic(tree, err);
1145                 cache_state(prealloc, cached_state);
1146                 prealloc = NULL;
1147                 goto out;
1148         }
1149         state = rb_entry(node, struct extent_state, rb_node);
1150 hit_next:
1151         last_start = state->start;
1152         last_end = state->end;
1153
1154         /*
1155          * | ---- desired range ---- |
1156          * | state |
1157          *
1158          * Just lock what we found and keep going
1159          */
1160         if (state->start == start && state->end <= end) {
1161                 set_state_bits(tree, state, &bits, NULL);
1162                 cache_state(state, cached_state);
1163                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1164                 if (last_end == (u64)-1)
1165                         goto out;
1166                 start = last_end + 1;
1167                 if (start < end && state && state->start == start &&
1168                     !need_resched())
1169                         goto hit_next;
1170                 goto search_again;
1171         }
1172
1173         /*
1174          *     | ---- desired range ---- |
1175          * | state |
1176          *   or
1177          * | ------------- state -------------- |
1178          *
1179          * We need to split the extent we found, and may flip bits on
1180          * second half.
1181          *
1182          * If the extent we found extends past our
1183          * range, we just split and search again.  It'll get split
1184          * again the next time though.
1185          *
1186          * If the extent we found is inside our range, we set the
1187          * desired bit on it.
1188          */
1189         if (state->start < start) {
1190                 prealloc = alloc_extent_state_atomic(prealloc);
1191                 if (!prealloc) {
1192                         err = -ENOMEM;
1193                         goto out;
1194                 }
1195                 err = split_state(tree, state, prealloc, start);
1196                 if (err)
1197                         extent_io_tree_panic(tree, err);
1198                 prealloc = NULL;
1199                 if (err)
1200                         goto out;
1201                 if (state->end <= end) {
1202                         set_state_bits(tree, state, &bits, NULL);
1203                         cache_state(state, cached_state);
1204                         state = clear_state_bit(tree, state, &clear_bits, 0,
1205                                                 NULL);
1206                         if (last_end == (u64)-1)
1207                                 goto out;
1208                         start = last_end + 1;
1209                         if (start < end && state && state->start == start &&
1210                             !need_resched())
1211                                 goto hit_next;
1212                 }
1213                 goto search_again;
1214         }
1215         /*
1216          * | ---- desired range ---- |
1217          *     | state | or               | state |
1218          *
1219          * There's a hole, we need to insert something in it and
1220          * ignore the extent we found.
1221          */
1222         if (state->start > start) {
1223                 u64 this_end;
1224                 if (end < last_start)
1225                         this_end = end;
1226                 else
1227                         this_end = last_start - 1;
1228
1229                 prealloc = alloc_extent_state_atomic(prealloc);
1230                 if (!prealloc) {
1231                         err = -ENOMEM;
1232                         goto out;
1233                 }
1234
1235                 /*
1236                  * Avoid to free 'prealloc' if it can be merged with
1237                  * the later extent.
1238                  */
1239                 err = insert_state(tree, prealloc, start, this_end,
1240                                    NULL, NULL, &bits, NULL);
1241                 if (err)
1242                         extent_io_tree_panic(tree, err);
1243                 cache_state(prealloc, cached_state);
1244                 prealloc = NULL;
1245                 start = this_end + 1;
1246                 goto search_again;
1247         }
1248         /*
1249          * | ---- desired range ---- |
1250          *                        | state |
1251          * We need to split the extent, and set the bit
1252          * on the first half
1253          */
1254         if (state->start <= end && state->end > end) {
1255                 prealloc = alloc_extent_state_atomic(prealloc);
1256                 if (!prealloc) {
1257                         err = -ENOMEM;
1258                         goto out;
1259                 }
1260
1261                 err = split_state(tree, state, prealloc, end + 1);
1262                 if (err)
1263                         extent_io_tree_panic(tree, err);
1264
1265                 set_state_bits(tree, prealloc, &bits, NULL);
1266                 cache_state(prealloc, cached_state);
1267                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1268                 prealloc = NULL;
1269                 goto out;
1270         }
1271
1272 search_again:
1273         if (start > end)
1274                 goto out;
1275         spin_unlock(&tree->lock);
1276         cond_resched();
1277         first_iteration = false;
1278         goto again;
1279
1280 out:
1281         spin_unlock(&tree->lock);
1282         if (prealloc)
1283                 free_extent_state(prealloc);
1284
1285         return err;
1286 }
1287
1288 /* wrappers around set/clear extent bit */
1289 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1290                            unsigned bits, struct extent_changeset *changeset)
1291 {
1292         /*
1293          * We don't support EXTENT_LOCKED yet, as current changeset will
1294          * record any bits changed, so for EXTENT_LOCKED case, it will
1295          * either fail with -EEXIST or changeset will record the whole
1296          * range.
1297          */
1298         BUG_ON(bits & EXTENT_LOCKED);
1299
1300         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1301                                 changeset);
1302 }
1303
1304 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1305                      unsigned bits, int wake, int delete,
1306                      struct extent_state **cached, gfp_t mask)
1307 {
1308         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1309                                   cached, mask, NULL);
1310 }
1311
1312 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1313                 unsigned bits, struct extent_changeset *changeset)
1314 {
1315         /*
1316          * Don't support EXTENT_LOCKED case, same reason as
1317          * set_record_extent_bits().
1318          */
1319         BUG_ON(bits & EXTENT_LOCKED);
1320
1321         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1322                                   changeset);
1323 }
1324
1325 /*
1326  * either insert or lock state struct between start and end use mask to tell
1327  * us if waiting is desired.
1328  */
1329 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1330                      struct extent_state **cached_state)
1331 {
1332         int err;
1333         u64 failed_start;
1334
1335         while (1) {
1336                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1337                                        EXTENT_LOCKED, &failed_start,
1338                                        cached_state, GFP_NOFS, NULL);
1339                 if (err == -EEXIST) {
1340                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1341                         start = failed_start;
1342                 } else
1343                         break;
1344                 WARN_ON(start > end);
1345         }
1346         return err;
1347 }
1348
1349 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1350 {
1351         int err;
1352         u64 failed_start;
1353
1354         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1355                                &failed_start, NULL, GFP_NOFS, NULL);
1356         if (err == -EEXIST) {
1357                 if (failed_start > start)
1358                         clear_extent_bit(tree, start, failed_start - 1,
1359                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1360                 return 0;
1361         }
1362         return 1;
1363 }
1364
1365 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1366 {
1367         unsigned long index = start >> PAGE_SHIFT;
1368         unsigned long end_index = end >> PAGE_SHIFT;
1369         struct page *page;
1370
1371         while (index <= end_index) {
1372                 page = find_get_page(inode->i_mapping, index);
1373                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1374                 clear_page_dirty_for_io(page);
1375                 put_page(page);
1376                 index++;
1377         }
1378 }
1379
1380 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1381 {
1382         unsigned long index = start >> PAGE_SHIFT;
1383         unsigned long end_index = end >> PAGE_SHIFT;
1384         struct page *page;
1385
1386         while (index <= end_index) {
1387                 page = find_get_page(inode->i_mapping, index);
1388                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1389                 __set_page_dirty_nobuffers(page);
1390                 account_page_redirty(page);
1391                 put_page(page);
1392                 index++;
1393         }
1394 }
1395
1396 /*
1397  * helper function to set both pages and extents in the tree writeback
1398  */
1399 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1400 {
1401         unsigned long index = start >> PAGE_SHIFT;
1402         unsigned long end_index = end >> PAGE_SHIFT;
1403         struct page *page;
1404
1405         while (index <= end_index) {
1406                 page = find_get_page(tree->mapping, index);
1407                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1408                 set_page_writeback(page);
1409                 put_page(page);
1410                 index++;
1411         }
1412 }
1413
1414 /* find the first state struct with 'bits' set after 'start', and
1415  * return it.  tree->lock must be held.  NULL will returned if
1416  * nothing was found after 'start'
1417  */
1418 static struct extent_state *
1419 find_first_extent_bit_state(struct extent_io_tree *tree,
1420                             u64 start, unsigned bits)
1421 {
1422         struct rb_node *node;
1423         struct extent_state *state;
1424
1425         /*
1426          * this search will find all the extents that end after
1427          * our range starts.
1428          */
1429         node = tree_search(tree, start);
1430         if (!node)
1431                 goto out;
1432
1433         while (1) {
1434                 state = rb_entry(node, struct extent_state, rb_node);
1435                 if (state->end >= start && (state->state & bits))
1436                         return state;
1437
1438                 node = rb_next(node);
1439                 if (!node)
1440                         break;
1441         }
1442 out:
1443         return NULL;
1444 }
1445
1446 /*
1447  * find the first offset in the io tree with 'bits' set. zero is
1448  * returned if we find something, and *start_ret and *end_ret are
1449  * set to reflect the state struct that was found.
1450  *
1451  * If nothing was found, 1 is returned. If found something, return 0.
1452  */
1453 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1454                           u64 *start_ret, u64 *end_ret, unsigned bits,
1455                           struct extent_state **cached_state)
1456 {
1457         struct extent_state *state;
1458         struct rb_node *n;
1459         int ret = 1;
1460
1461         spin_lock(&tree->lock);
1462         if (cached_state && *cached_state) {
1463                 state = *cached_state;
1464                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1465                         n = rb_next(&state->rb_node);
1466                         while (n) {
1467                                 state = rb_entry(n, struct extent_state,
1468                                                  rb_node);
1469                                 if (state->state & bits)
1470                                         goto got_it;
1471                                 n = rb_next(n);
1472                         }
1473                         free_extent_state(*cached_state);
1474                         *cached_state = NULL;
1475                         goto out;
1476                 }
1477                 free_extent_state(*cached_state);
1478                 *cached_state = NULL;
1479         }
1480
1481         state = find_first_extent_bit_state(tree, start, bits);
1482 got_it:
1483         if (state) {
1484                 cache_state_if_flags(state, cached_state, 0);
1485                 *start_ret = state->start;
1486                 *end_ret = state->end;
1487                 ret = 0;
1488         }
1489 out:
1490         spin_unlock(&tree->lock);
1491         return ret;
1492 }
1493
1494 /*
1495  * find a contiguous range of bytes in the file marked as delalloc, not
1496  * more than 'max_bytes'.  start and end are used to return the range,
1497  *
1498  * 1 is returned if we find something, 0 if nothing was in the tree
1499  */
1500 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1501                                         u64 *start, u64 *end, u64 max_bytes,
1502                                         struct extent_state **cached_state)
1503 {
1504         struct rb_node *node;
1505         struct extent_state *state;
1506         u64 cur_start = *start;
1507         u64 found = 0;
1508         u64 total_bytes = 0;
1509
1510         spin_lock(&tree->lock);
1511
1512         /*
1513          * this search will find all the extents that end after
1514          * our range starts.
1515          */
1516         node = tree_search(tree, cur_start);
1517         if (!node) {
1518                 if (!found)
1519                         *end = (u64)-1;
1520                 goto out;
1521         }
1522
1523         while (1) {
1524                 state = rb_entry(node, struct extent_state, rb_node);
1525                 if (found && (state->start != cur_start ||
1526                               (state->state & EXTENT_BOUNDARY))) {
1527                         goto out;
1528                 }
1529                 if (!(state->state & EXTENT_DELALLOC)) {
1530                         if (!found)
1531                                 *end = state->end;
1532                         goto out;
1533                 }
1534                 if (!found) {
1535                         *start = state->start;
1536                         *cached_state = state;
1537                         atomic_inc(&state->refs);
1538                 }
1539                 found++;
1540                 *end = state->end;
1541                 cur_start = state->end + 1;
1542                 node = rb_next(node);
1543                 total_bytes += state->end - state->start + 1;
1544                 if (total_bytes >= max_bytes)
1545                         break;
1546                 if (!node)
1547                         break;
1548         }
1549 out:
1550         spin_unlock(&tree->lock);
1551         return found;
1552 }
1553
1554 static noinline void __unlock_for_delalloc(struct inode *inode,
1555                                            struct page *locked_page,
1556                                            u64 start, u64 end)
1557 {
1558         int ret;
1559         struct page *pages[16];
1560         unsigned long index = start >> PAGE_SHIFT;
1561         unsigned long end_index = end >> PAGE_SHIFT;
1562         unsigned long nr_pages = end_index - index + 1;
1563         int i;
1564
1565         if (index == locked_page->index && end_index == index)
1566                 return;
1567
1568         while (nr_pages > 0) {
1569                 ret = find_get_pages_contig(inode->i_mapping, index,
1570                                      min_t(unsigned long, nr_pages,
1571                                      ARRAY_SIZE(pages)), pages);
1572                 for (i = 0; i < ret; i++) {
1573                         if (pages[i] != locked_page)
1574                                 unlock_page(pages[i]);
1575                         put_page(pages[i]);
1576                 }
1577                 nr_pages -= ret;
1578                 index += ret;
1579                 cond_resched();
1580         }
1581 }
1582
1583 static noinline int lock_delalloc_pages(struct inode *inode,
1584                                         struct page *locked_page,
1585                                         u64 delalloc_start,
1586                                         u64 delalloc_end)
1587 {
1588         unsigned long index = delalloc_start >> PAGE_SHIFT;
1589         unsigned long start_index = index;
1590         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1591         unsigned long pages_locked = 0;
1592         struct page *pages[16];
1593         unsigned long nrpages;
1594         int ret;
1595         int i;
1596
1597         /* the caller is responsible for locking the start index */
1598         if (index == locked_page->index && index == end_index)
1599                 return 0;
1600
1601         /* skip the page at the start index */
1602         nrpages = end_index - index + 1;
1603         while (nrpages > 0) {
1604                 ret = find_get_pages_contig(inode->i_mapping, index,
1605                                      min_t(unsigned long,
1606                                      nrpages, ARRAY_SIZE(pages)), pages);
1607                 if (ret == 0) {
1608                         ret = -EAGAIN;
1609                         goto done;
1610                 }
1611                 /* now we have an array of pages, lock them all */
1612                 for (i = 0; i < ret; i++) {
1613                         /*
1614                          * the caller is taking responsibility for
1615                          * locked_page
1616                          */
1617                         if (pages[i] != locked_page) {
1618                                 lock_page(pages[i]);
1619                                 if (!PageDirty(pages[i]) ||
1620                                     pages[i]->mapping != inode->i_mapping) {
1621                                         ret = -EAGAIN;
1622                                         unlock_page(pages[i]);
1623                                         put_page(pages[i]);
1624                                         goto done;
1625                                 }
1626                         }
1627                         put_page(pages[i]);
1628                         pages_locked++;
1629                 }
1630                 nrpages -= ret;
1631                 index += ret;
1632                 cond_resched();
1633         }
1634         ret = 0;
1635 done:
1636         if (ret && pages_locked) {
1637                 __unlock_for_delalloc(inode, locked_page,
1638                               delalloc_start,
1639                               ((u64)(start_index + pages_locked - 1)) <<
1640                               PAGE_SHIFT);
1641         }
1642         return ret;
1643 }
1644
1645 /*
1646  * find a contiguous range of bytes in the file marked as delalloc, not
1647  * more than 'max_bytes'.  start and end are used to return the range,
1648  *
1649  * 1 is returned if we find something, 0 if nothing was in the tree
1650  */
1651 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1652                                     struct extent_io_tree *tree,
1653                                     struct page *locked_page, u64 *start,
1654                                     u64 *end, u64 max_bytes)
1655 {
1656         u64 delalloc_start;
1657         u64 delalloc_end;
1658         u64 found;
1659         struct extent_state *cached_state = NULL;
1660         int ret;
1661         int loops = 0;
1662
1663 again:
1664         /* step one, find a bunch of delalloc bytes starting at start */
1665         delalloc_start = *start;
1666         delalloc_end = 0;
1667         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1668                                     max_bytes, &cached_state);
1669         if (!found || delalloc_end <= *start) {
1670                 *start = delalloc_start;
1671                 *end = delalloc_end;
1672                 free_extent_state(cached_state);
1673                 return 0;
1674         }
1675
1676         /*
1677          * start comes from the offset of locked_page.  We have to lock
1678          * pages in order, so we can't process delalloc bytes before
1679          * locked_page
1680          */
1681         if (delalloc_start < *start)
1682                 delalloc_start = *start;
1683
1684         /*
1685          * make sure to limit the number of pages we try to lock down
1686          */
1687         if (delalloc_end + 1 - delalloc_start > max_bytes)
1688                 delalloc_end = delalloc_start + max_bytes - 1;
1689
1690         /* step two, lock all the pages after the page that has start */
1691         ret = lock_delalloc_pages(inode, locked_page,
1692                                   delalloc_start, delalloc_end);
1693         if (ret == -EAGAIN) {
1694                 /* some of the pages are gone, lets avoid looping by
1695                  * shortening the size of the delalloc range we're searching
1696                  */
1697                 free_extent_state(cached_state);
1698                 cached_state = NULL;
1699                 if (!loops) {
1700                         max_bytes = PAGE_SIZE;
1701                         loops = 1;
1702                         goto again;
1703                 } else {
1704                         found = 0;
1705                         goto out_failed;
1706                 }
1707         }
1708         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1709
1710         /* step three, lock the state bits for the whole range */
1711         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1712
1713         /* then test to make sure it is all still delalloc */
1714         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1715                              EXTENT_DELALLOC, 1, cached_state);
1716         if (!ret) {
1717                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1718                                      &cached_state, GFP_NOFS);
1719                 __unlock_for_delalloc(inode, locked_page,
1720                               delalloc_start, delalloc_end);
1721                 cond_resched();
1722                 goto again;
1723         }
1724         free_extent_state(cached_state);
1725         *start = delalloc_start;
1726         *end = delalloc_end;
1727 out_failed:
1728         return found;
1729 }
1730
1731 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1732                                  struct page *locked_page,
1733                                  unsigned clear_bits,
1734                                  unsigned long page_ops)
1735 {
1736         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1737         int ret;
1738         struct page *pages[16];
1739         unsigned long index = start >> PAGE_SHIFT;
1740         unsigned long end_index = end >> PAGE_SHIFT;
1741         unsigned long nr_pages = end_index - index + 1;
1742         int i;
1743
1744         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1745         if (page_ops == 0)
1746                 return;
1747
1748         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1749                 mapping_set_error(inode->i_mapping, -EIO);
1750
1751         while (nr_pages > 0) {
1752                 ret = find_get_pages_contig(inode->i_mapping, index,
1753                                      min_t(unsigned long,
1754                                      nr_pages, ARRAY_SIZE(pages)), pages);
1755                 for (i = 0; i < ret; i++) {
1756
1757                         if (page_ops & PAGE_SET_PRIVATE2)
1758                                 SetPagePrivate2(pages[i]);
1759
1760                         if (pages[i] == locked_page) {
1761                                 put_page(pages[i]);
1762                                 continue;
1763                         }
1764                         if (page_ops & PAGE_CLEAR_DIRTY)
1765                                 clear_page_dirty_for_io(pages[i]);
1766                         if (page_ops & PAGE_SET_WRITEBACK)
1767                                 set_page_writeback(pages[i]);
1768                         if (page_ops & PAGE_SET_ERROR)
1769                                 SetPageError(pages[i]);
1770                         if (page_ops & PAGE_END_WRITEBACK)
1771                                 end_page_writeback(pages[i]);
1772                         if (page_ops & PAGE_UNLOCK)
1773                                 unlock_page(pages[i]);
1774                         put_page(pages[i]);
1775                 }
1776                 nr_pages -= ret;
1777                 index += ret;
1778                 cond_resched();
1779         }
1780 }
1781
1782 /*
1783  * count the number of bytes in the tree that have a given bit(s)
1784  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1785  * cached.  The total number found is returned.
1786  */
1787 u64 count_range_bits(struct extent_io_tree *tree,
1788                      u64 *start, u64 search_end, u64 max_bytes,
1789                      unsigned bits, int contig)
1790 {
1791         struct rb_node *node;
1792         struct extent_state *state;
1793         u64 cur_start = *start;
1794         u64 total_bytes = 0;
1795         u64 last = 0;
1796         int found = 0;
1797
1798         if (WARN_ON(search_end <= cur_start))
1799                 return 0;
1800
1801         spin_lock(&tree->lock);
1802         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1803                 total_bytes = tree->dirty_bytes;
1804                 goto out;
1805         }
1806         /*
1807          * this search will find all the extents that end after
1808          * our range starts.
1809          */
1810         node = tree_search(tree, cur_start);
1811         if (!node)
1812                 goto out;
1813
1814         while (1) {
1815                 state = rb_entry(node, struct extent_state, rb_node);
1816                 if (state->start > search_end)
1817                         break;
1818                 if (contig && found && state->start > last + 1)
1819                         break;
1820                 if (state->end >= cur_start && (state->state & bits) == bits) {
1821                         total_bytes += min(search_end, state->end) + 1 -
1822                                        max(cur_start, state->start);
1823                         if (total_bytes >= max_bytes)
1824                                 break;
1825                         if (!found) {
1826                                 *start = max(cur_start, state->start);
1827                                 found = 1;
1828                         }
1829                         last = state->end;
1830                 } else if (contig && found) {
1831                         break;
1832                 }
1833                 node = rb_next(node);
1834                 if (!node)
1835                         break;
1836         }
1837 out:
1838         spin_unlock(&tree->lock);
1839         return total_bytes;
1840 }
1841
1842 /*
1843  * set the private field for a given byte offset in the tree.  If there isn't
1844  * an extent_state there already, this does nothing.
1845  */
1846 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1847                 struct io_failure_record *failrec)
1848 {
1849         struct rb_node *node;
1850         struct extent_state *state;
1851         int ret = 0;
1852
1853         spin_lock(&tree->lock);
1854         /*
1855          * this search will find all the extents that end after
1856          * our range starts.
1857          */
1858         node = tree_search(tree, start);
1859         if (!node) {
1860                 ret = -ENOENT;
1861                 goto out;
1862         }
1863         state = rb_entry(node, struct extent_state, rb_node);
1864         if (state->start != start) {
1865                 ret = -ENOENT;
1866                 goto out;
1867         }
1868         state->failrec = failrec;
1869 out:
1870         spin_unlock(&tree->lock);
1871         return ret;
1872 }
1873
1874 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1875                 struct io_failure_record **failrec)
1876 {
1877         struct rb_node *node;
1878         struct extent_state *state;
1879         int ret = 0;
1880
1881         spin_lock(&tree->lock);
1882         /*
1883          * this search will find all the extents that end after
1884          * our range starts.
1885          */
1886         node = tree_search(tree, start);
1887         if (!node) {
1888                 ret = -ENOENT;
1889                 goto out;
1890         }
1891         state = rb_entry(node, struct extent_state, rb_node);
1892         if (state->start != start) {
1893                 ret = -ENOENT;
1894                 goto out;
1895         }
1896         *failrec = state->failrec;
1897 out:
1898         spin_unlock(&tree->lock);
1899         return ret;
1900 }
1901
1902 /*
1903  * searches a range in the state tree for a given mask.
1904  * If 'filled' == 1, this returns 1 only if every extent in the tree
1905  * has the bits set.  Otherwise, 1 is returned if any bit in the
1906  * range is found set.
1907  */
1908 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1909                    unsigned bits, int filled, struct extent_state *cached)
1910 {
1911         struct extent_state *state = NULL;
1912         struct rb_node *node;
1913         int bitset = 0;
1914
1915         spin_lock(&tree->lock);
1916         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1917             cached->end > start)
1918                 node = &cached->rb_node;
1919         else
1920                 node = tree_search(tree, start);
1921         while (node && start <= end) {
1922                 state = rb_entry(node, struct extent_state, rb_node);
1923
1924                 if (filled && state->start > start) {
1925                         bitset = 0;
1926                         break;
1927                 }
1928
1929                 if (state->start > end)
1930                         break;
1931
1932                 if (state->state & bits) {
1933                         bitset = 1;
1934                         if (!filled)
1935                                 break;
1936                 } else if (filled) {
1937                         bitset = 0;
1938                         break;
1939                 }
1940
1941                 if (state->end == (u64)-1)
1942                         break;
1943
1944                 start = state->end + 1;
1945                 if (start > end)
1946                         break;
1947                 node = rb_next(node);
1948                 if (!node) {
1949                         if (filled)
1950                                 bitset = 0;
1951                         break;
1952                 }
1953         }
1954         spin_unlock(&tree->lock);
1955         return bitset;
1956 }
1957
1958 /*
1959  * helper function to set a given page up to date if all the
1960  * extents in the tree for that page are up to date
1961  */
1962 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1963 {
1964         u64 start = page_offset(page);
1965         u64 end = start + PAGE_SIZE - 1;
1966         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1967                 SetPageUptodate(page);
1968 }
1969
1970 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1971 {
1972         int ret;
1973         int err = 0;
1974         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1975
1976         set_state_failrec(failure_tree, rec->start, NULL);
1977         ret = clear_extent_bits(failure_tree, rec->start,
1978                                 rec->start + rec->len - 1,
1979                                 EXTENT_LOCKED | EXTENT_DIRTY);
1980         if (ret)
1981                 err = ret;
1982
1983         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1984                                 rec->start + rec->len - 1,
1985                                 EXTENT_DAMAGED);
1986         if (ret && !err)
1987                 err = ret;
1988
1989         kfree(rec);
1990         return err;
1991 }
1992
1993 /*
1994  * this bypasses the standard btrfs submit functions deliberately, as
1995  * the standard behavior is to write all copies in a raid setup. here we only
1996  * want to write the one bad copy. so we do the mapping for ourselves and issue
1997  * submit_bio directly.
1998  * to avoid any synchronization issues, wait for the data after writing, which
1999  * actually prevents the read that triggered the error from finishing.
2000  * currently, there can be no more than two copies of every data bit. thus,
2001  * exactly one rewrite is required.
2002  */
2003 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2004                       struct page *page, unsigned int pg_offset, int mirror_num)
2005 {
2006         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2007         struct bio *bio;
2008         struct btrfs_device *dev;
2009         u64 map_length = 0;
2010         u64 sector;
2011         struct btrfs_bio *bbio = NULL;
2012         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2013         int ret;
2014
2015         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2016         BUG_ON(!mirror_num);
2017
2018         /* we can't repair anything in raid56 yet */
2019         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2020                 return 0;
2021
2022         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2023         if (!bio)
2024                 return -EIO;
2025         bio->bi_iter.bi_size = 0;
2026         map_length = length;
2027
2028         /*
2029          * Avoid races with device replace and make sure our bbio has devices
2030          * associated to its stripes that don't go away while we are doing the
2031          * read repair operation.
2032          */
2033         btrfs_bio_counter_inc_blocked(fs_info);
2034         ret = btrfs_map_block(fs_info, WRITE, logical,
2035                               &map_length, &bbio, mirror_num);
2036         if (ret) {
2037                 btrfs_bio_counter_dec(fs_info);
2038                 bio_put(bio);
2039                 return -EIO;
2040         }
2041         BUG_ON(mirror_num != bbio->mirror_num);
2042         sector = bbio->stripes[mirror_num-1].physical >> 9;
2043         bio->bi_iter.bi_sector = sector;
2044         dev = bbio->stripes[mirror_num-1].dev;
2045         btrfs_put_bbio(bbio);
2046         if (!dev || !dev->bdev || !dev->writeable) {
2047                 btrfs_bio_counter_dec(fs_info);
2048                 bio_put(bio);
2049                 return -EIO;
2050         }
2051         bio->bi_bdev = dev->bdev;
2052         bio_set_op_attrs(bio, REQ_OP_WRITE, WRITE_SYNC);
2053         bio_add_page(bio, page, length, pg_offset);
2054
2055         if (btrfsic_submit_bio_wait(bio)) {
2056                 /* try to remap that extent elsewhere? */
2057                 btrfs_bio_counter_dec(fs_info);
2058                 bio_put(bio);
2059                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2060                 return -EIO;
2061         }
2062
2063         btrfs_info_rl_in_rcu(fs_info,
2064                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2065                                   btrfs_ino(inode), start,
2066                                   rcu_str_deref(dev->name), sector);
2067         btrfs_bio_counter_dec(fs_info);
2068         bio_put(bio);
2069         return 0;
2070 }
2071
2072 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2073                          int mirror_num)
2074 {
2075         u64 start = eb->start;
2076         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2077         int ret = 0;
2078
2079         if (root->fs_info->sb->s_flags & MS_RDONLY)
2080                 return -EROFS;
2081
2082         for (i = 0; i < num_pages; i++) {
2083                 struct page *p = eb->pages[i];
2084
2085                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2086                                         PAGE_SIZE, start, p,
2087                                         start - page_offset(p), mirror_num);
2088                 if (ret)
2089                         break;
2090                 start += PAGE_SIZE;
2091         }
2092
2093         return ret;
2094 }
2095
2096 /*
2097  * each time an IO finishes, we do a fast check in the IO failure tree
2098  * to see if we need to process or clean up an io_failure_record
2099  */
2100 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2101                      unsigned int pg_offset)
2102 {
2103         u64 private;
2104         struct io_failure_record *failrec;
2105         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2106         struct extent_state *state;
2107         int num_copies;
2108         int ret;
2109
2110         private = 0;
2111         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2112                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2113         if (!ret)
2114                 return 0;
2115
2116         ret = get_state_failrec(&BTRFS_I(inode)->io_failure_tree, start,
2117                         &failrec);
2118         if (ret)
2119                 return 0;
2120
2121         BUG_ON(!failrec->this_mirror);
2122
2123         if (failrec->in_validation) {
2124                 /* there was no real error, just free the record */
2125                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2126                          failrec->start);
2127                 goto out;
2128         }
2129         if (fs_info->sb->s_flags & MS_RDONLY)
2130                 goto out;
2131
2132         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2133         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2134                                             failrec->start,
2135                                             EXTENT_LOCKED);
2136         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2137
2138         if (state && state->start <= failrec->start &&
2139             state->end >= failrec->start + failrec->len - 1) {
2140                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2141                                               failrec->len);
2142                 if (num_copies > 1)  {
2143                         repair_io_failure(inode, start, failrec->len,
2144                                           failrec->logical, page,
2145                                           pg_offset, failrec->failed_mirror);
2146                 }
2147         }
2148
2149 out:
2150         free_io_failure(inode, failrec);
2151
2152         return 0;
2153 }
2154
2155 /*
2156  * Can be called when
2157  * - hold extent lock
2158  * - under ordered extent
2159  * - the inode is freeing
2160  */
2161 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2162 {
2163         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2164         struct io_failure_record *failrec;
2165         struct extent_state *state, *next;
2166
2167         if (RB_EMPTY_ROOT(&failure_tree->state))
2168                 return;
2169
2170         spin_lock(&failure_tree->lock);
2171         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2172         while (state) {
2173                 if (state->start > end)
2174                         break;
2175
2176                 ASSERT(state->end <= end);
2177
2178                 next = next_state(state);
2179
2180                 failrec = state->failrec;
2181                 free_extent_state(state);
2182                 kfree(failrec);
2183
2184                 state = next;
2185         }
2186         spin_unlock(&failure_tree->lock);
2187 }
2188
2189 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2190                 struct io_failure_record **failrec_ret)
2191 {
2192         struct io_failure_record *failrec;
2193         struct extent_map *em;
2194         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2195         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2196         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2197         int ret;
2198         u64 logical;
2199
2200         ret = get_state_failrec(failure_tree, start, &failrec);
2201         if (ret) {
2202                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2203                 if (!failrec)
2204                         return -ENOMEM;
2205
2206                 failrec->start = start;
2207                 failrec->len = end - start + 1;
2208                 failrec->this_mirror = 0;
2209                 failrec->bio_flags = 0;
2210                 failrec->in_validation = 0;
2211
2212                 read_lock(&em_tree->lock);
2213                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2214                 if (!em) {
2215                         read_unlock(&em_tree->lock);
2216                         kfree(failrec);
2217                         return -EIO;
2218                 }
2219
2220                 if (em->start > start || em->start + em->len <= start) {
2221                         free_extent_map(em);
2222                         em = NULL;
2223                 }
2224                 read_unlock(&em_tree->lock);
2225                 if (!em) {
2226                         kfree(failrec);
2227                         return -EIO;
2228                 }
2229
2230                 logical = start - em->start;
2231                 logical = em->block_start + logical;
2232                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2233                         logical = em->block_start;
2234                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2235                         extent_set_compress_type(&failrec->bio_flags,
2236                                                  em->compress_type);
2237                 }
2238
2239                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2240                          logical, start, failrec->len);
2241
2242                 failrec->logical = logical;
2243                 free_extent_map(em);
2244
2245                 /* set the bits in the private failure tree */
2246                 ret = set_extent_bits(failure_tree, start, end,
2247                                         EXTENT_LOCKED | EXTENT_DIRTY);
2248                 if (ret >= 0)
2249                         ret = set_state_failrec(failure_tree, start, failrec);
2250                 /* set the bits in the inode's tree */
2251                 if (ret >= 0)
2252                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2253                 if (ret < 0) {
2254                         kfree(failrec);
2255                         return ret;
2256                 }
2257         } else {
2258                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2259                          failrec->logical, failrec->start, failrec->len,
2260                          failrec->in_validation);
2261                 /*
2262                  * when data can be on disk more than twice, add to failrec here
2263                  * (e.g. with a list for failed_mirror) to make
2264                  * clean_io_failure() clean all those errors at once.
2265                  */
2266         }
2267
2268         *failrec_ret = failrec;
2269
2270         return 0;
2271 }
2272
2273 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2274                            struct io_failure_record *failrec, int failed_mirror)
2275 {
2276         int num_copies;
2277
2278         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2279                                       failrec->logical, failrec->len);
2280         if (num_copies == 1) {
2281                 /*
2282                  * we only have a single copy of the data, so don't bother with
2283                  * all the retry and error correction code that follows. no
2284                  * matter what the error is, it is very likely to persist.
2285                  */
2286                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2287                          num_copies, failrec->this_mirror, failed_mirror);
2288                 return 0;
2289         }
2290
2291         /*
2292          * there are two premises:
2293          *      a) deliver good data to the caller
2294          *      b) correct the bad sectors on disk
2295          */
2296         if (failed_bio->bi_vcnt > 1) {
2297                 /*
2298                  * to fulfill b), we need to know the exact failing sectors, as
2299                  * we don't want to rewrite any more than the failed ones. thus,
2300                  * we need separate read requests for the failed bio
2301                  *
2302                  * if the following BUG_ON triggers, our validation request got
2303                  * merged. we need separate requests for our algorithm to work.
2304                  */
2305                 BUG_ON(failrec->in_validation);
2306                 failrec->in_validation = 1;
2307                 failrec->this_mirror = failed_mirror;
2308         } else {
2309                 /*
2310                  * we're ready to fulfill a) and b) alongside. get a good copy
2311                  * of the failed sector and if we succeed, we have setup
2312                  * everything for repair_io_failure to do the rest for us.
2313                  */
2314                 if (failrec->in_validation) {
2315                         BUG_ON(failrec->this_mirror != failed_mirror);
2316                         failrec->in_validation = 0;
2317                         failrec->this_mirror = 0;
2318                 }
2319                 failrec->failed_mirror = failed_mirror;
2320                 failrec->this_mirror++;
2321                 if (failrec->this_mirror == failed_mirror)
2322                         failrec->this_mirror++;
2323         }
2324
2325         if (failrec->this_mirror > num_copies) {
2326                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2327                          num_copies, failrec->this_mirror, failed_mirror);
2328                 return 0;
2329         }
2330
2331         return 1;
2332 }
2333
2334
2335 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2336                                     struct io_failure_record *failrec,
2337                                     struct page *page, int pg_offset, int icsum,
2338                                     bio_end_io_t *endio_func, void *data)
2339 {
2340         struct bio *bio;
2341         struct btrfs_io_bio *btrfs_failed_bio;
2342         struct btrfs_io_bio *btrfs_bio;
2343
2344         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2345         if (!bio)
2346                 return NULL;
2347
2348         bio->bi_end_io = endio_func;
2349         bio->bi_iter.bi_sector = failrec->logical >> 9;
2350         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2351         bio->bi_iter.bi_size = 0;
2352         bio->bi_private = data;
2353
2354         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2355         if (btrfs_failed_bio->csum) {
2356                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2357                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2358
2359                 btrfs_bio = btrfs_io_bio(bio);
2360                 btrfs_bio->csum = btrfs_bio->csum_inline;
2361                 icsum *= csum_size;
2362                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2363                        csum_size);
2364         }
2365
2366         bio_add_page(bio, page, failrec->len, pg_offset);
2367
2368         return bio;
2369 }
2370
2371 /*
2372  * this is a generic handler for readpage errors (default
2373  * readpage_io_failed_hook). if other copies exist, read those and write back
2374  * good data to the failed position. does not investigate in remapping the
2375  * failed extent elsewhere, hoping the device will be smart enough to do this as
2376  * needed
2377  */
2378
2379 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2380                               struct page *page, u64 start, u64 end,
2381                               int failed_mirror)
2382 {
2383         struct io_failure_record *failrec;
2384         struct inode *inode = page->mapping->host;
2385         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2386         struct bio *bio;
2387         int read_mode;
2388         int ret;
2389
2390         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2391
2392         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2393         if (ret)
2394                 return ret;
2395
2396         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2397         if (!ret) {
2398                 free_io_failure(inode, failrec);
2399                 return -EIO;
2400         }
2401
2402         if (failed_bio->bi_vcnt > 1)
2403                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2404         else
2405                 read_mode = READ_SYNC;
2406
2407         phy_offset >>= inode->i_sb->s_blocksize_bits;
2408         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2409                                       start - page_offset(page),
2410                                       (int)phy_offset, failed_bio->bi_end_io,
2411                                       NULL);
2412         if (!bio) {
2413                 free_io_failure(inode, failrec);
2414                 return -EIO;
2415         }
2416         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2417
2418         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2419                  read_mode, failrec->this_mirror, failrec->in_validation);
2420
2421         ret = tree->ops->submit_bio_hook(inode, bio, failrec->this_mirror,
2422                                          failrec->bio_flags, 0);
2423         if (ret) {
2424                 free_io_failure(inode, failrec);
2425                 bio_put(bio);
2426         }
2427
2428         return ret;
2429 }
2430
2431 /* lots and lots of room for performance fixes in the end_bio funcs */
2432
2433 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2434 {
2435         int uptodate = (err == 0);
2436         struct extent_io_tree *tree;
2437         int ret = 0;
2438
2439         tree = &BTRFS_I(page->mapping->host)->io_tree;
2440
2441         if (tree->ops && tree->ops->writepage_end_io_hook) {
2442                 ret = tree->ops->writepage_end_io_hook(page, start,
2443                                                end, NULL, uptodate);
2444                 if (ret)
2445                         uptodate = 0;
2446         }
2447
2448         if (!uptodate) {
2449                 ClearPageUptodate(page);
2450                 SetPageError(page);
2451                 ret = ret < 0 ? ret : -EIO;
2452                 mapping_set_error(page->mapping, ret);
2453         }
2454 }
2455
2456 /*
2457  * after a writepage IO is done, we need to:
2458  * clear the uptodate bits on error
2459  * clear the writeback bits in the extent tree for this IO
2460  * end_page_writeback if the page has no more pending IO
2461  *
2462  * Scheduling is not allowed, so the extent state tree is expected
2463  * to have one and only one object corresponding to this IO.
2464  */
2465 static void end_bio_extent_writepage(struct bio *bio)
2466 {
2467         struct bio_vec *bvec;
2468         u64 start;
2469         u64 end;
2470         int i;
2471
2472         bio_for_each_segment_all(bvec, bio, i) {
2473                 struct page *page = bvec->bv_page;
2474
2475                 /* We always issue full-page reads, but if some block
2476                  * in a page fails to read, blk_update_request() will
2477                  * advance bv_offset and adjust bv_len to compensate.
2478                  * Print a warning for nonzero offsets, and an error
2479                  * if they don't add up to a full page.  */
2480                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2481                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2482                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2483                                    "partial page write in btrfs with offset %u and length %u",
2484                                         bvec->bv_offset, bvec->bv_len);
2485                         else
2486                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2487                                    "incomplete page write in btrfs with offset %u and "
2488                                    "length %u",
2489                                         bvec->bv_offset, bvec->bv_len);
2490                 }
2491
2492                 start = page_offset(page);
2493                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2494
2495                 end_extent_writepage(page, bio->bi_error, start, end);
2496                 end_page_writeback(page);
2497         }
2498
2499         bio_put(bio);
2500 }
2501
2502 static void
2503 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2504                               int uptodate)
2505 {
2506         struct extent_state *cached = NULL;
2507         u64 end = start + len - 1;
2508
2509         if (uptodate && tree->track_uptodate)
2510                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2511         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2512 }
2513
2514 /*
2515  * after a readpage IO is done, we need to:
2516  * clear the uptodate bits on error
2517  * set the uptodate bits if things worked
2518  * set the page up to date if all extents in the tree are uptodate
2519  * clear the lock bit in the extent tree
2520  * unlock the page if there are no other extents locked for it
2521  *
2522  * Scheduling is not allowed, so the extent state tree is expected
2523  * to have one and only one object corresponding to this IO.
2524  */
2525 static void end_bio_extent_readpage(struct bio *bio)
2526 {
2527         struct bio_vec *bvec;
2528         int uptodate = !bio->bi_error;
2529         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2530         struct extent_io_tree *tree;
2531         u64 offset = 0;
2532         u64 start;
2533         u64 end;
2534         u64 len;
2535         u64 extent_start = 0;
2536         u64 extent_len = 0;
2537         int mirror;
2538         int ret;
2539         int i;
2540
2541         bio_for_each_segment_all(bvec, bio, i) {
2542                 struct page *page = bvec->bv_page;
2543                 struct inode *inode = page->mapping->host;
2544
2545                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2546                          "mirror=%u\n", (u64)bio->bi_iter.bi_sector,
2547                          bio->bi_error, io_bio->mirror_num);
2548                 tree = &BTRFS_I(inode)->io_tree;
2549
2550                 /* We always issue full-page reads, but if some block
2551                  * in a page fails to read, blk_update_request() will
2552                  * advance bv_offset and adjust bv_len to compensate.
2553                  * Print a warning for nonzero offsets, and an error
2554                  * if they don't add up to a full page.  */
2555                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2556                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2557                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2558                                    "partial page read in btrfs with offset %u and length %u",
2559                                         bvec->bv_offset, bvec->bv_len);
2560                         else
2561                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2562                                    "incomplete page read in btrfs with offset %u and "
2563                                    "length %u",
2564                                         bvec->bv_offset, bvec->bv_len);
2565                 }
2566
2567                 start = page_offset(page);
2568                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2569                 len = bvec->bv_len;
2570
2571                 mirror = io_bio->mirror_num;
2572                 if (likely(uptodate && tree->ops &&
2573                            tree->ops->readpage_end_io_hook)) {
2574                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2575                                                               page, start, end,
2576                                                               mirror);
2577                         if (ret)
2578                                 uptodate = 0;
2579                         else
2580                                 clean_io_failure(inode, start, page, 0);
2581                 }
2582
2583                 if (likely(uptodate))
2584                         goto readpage_ok;
2585
2586                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2587                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2588                         if (!ret && !bio->bi_error)
2589                                 uptodate = 1;
2590                 } else {
2591                         /*
2592                          * The generic bio_readpage_error handles errors the
2593                          * following way: If possible, new read requests are
2594                          * created and submitted and will end up in
2595                          * end_bio_extent_readpage as well (if we're lucky, not
2596                          * in the !uptodate case). In that case it returns 0 and
2597                          * we just go on with the next page in our bio. If it
2598                          * can't handle the error it will return -EIO and we
2599                          * remain responsible for that page.
2600                          */
2601                         ret = bio_readpage_error(bio, offset, page, start, end,
2602                                                  mirror);
2603                         if (ret == 0) {
2604                                 uptodate = !bio->bi_error;
2605                                 offset += len;
2606                                 continue;
2607                         }
2608                 }
2609 readpage_ok:
2610                 if (likely(uptodate)) {
2611                         loff_t i_size = i_size_read(inode);
2612                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2613                         unsigned off;
2614
2615                         /* Zero out the end if this page straddles i_size */
2616                         off = i_size & (PAGE_SIZE-1);
2617                         if (page->index == end_index && off)
2618                                 zero_user_segment(page, off, PAGE_SIZE);
2619                         SetPageUptodate(page);
2620                 } else {
2621                         ClearPageUptodate(page);
2622                         SetPageError(page);
2623                 }
2624                 unlock_page(page);
2625                 offset += len;
2626
2627                 if (unlikely(!uptodate)) {
2628                         if (extent_len) {
2629                                 endio_readpage_release_extent(tree,
2630                                                               extent_start,
2631                                                               extent_len, 1);
2632                                 extent_start = 0;
2633                                 extent_len = 0;
2634                         }
2635                         endio_readpage_release_extent(tree, start,
2636                                                       end - start + 1, 0);
2637                 } else if (!extent_len) {
2638                         extent_start = start;
2639                         extent_len = end + 1 - start;
2640                 } else if (extent_start + extent_len == start) {
2641                         extent_len += end + 1 - start;
2642                 } else {
2643                         endio_readpage_release_extent(tree, extent_start,
2644                                                       extent_len, uptodate);
2645                         extent_start = start;
2646                         extent_len = end + 1 - start;
2647                 }
2648         }
2649
2650         if (extent_len)
2651                 endio_readpage_release_extent(tree, extent_start, extent_len,
2652                                               uptodate);
2653         if (io_bio->end_io)
2654                 io_bio->end_io(io_bio, bio->bi_error);
2655         bio_put(bio);
2656 }
2657
2658 /*
2659  * this allocates from the btrfs_bioset.  We're returning a bio right now
2660  * but you can call btrfs_io_bio for the appropriate container_of magic
2661  */
2662 struct bio *
2663 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2664                 gfp_t gfp_flags)
2665 {
2666         struct btrfs_io_bio *btrfs_bio;
2667         struct bio *bio;
2668
2669         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2670
2671         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2672                 while (!bio && (nr_vecs /= 2)) {
2673                         bio = bio_alloc_bioset(gfp_flags,
2674                                                nr_vecs, btrfs_bioset);
2675                 }
2676         }
2677
2678         if (bio) {
2679                 bio->bi_bdev = bdev;
2680                 bio->bi_iter.bi_sector = first_sector;
2681                 btrfs_bio = btrfs_io_bio(bio);
2682                 btrfs_bio->csum = NULL;
2683                 btrfs_bio->csum_allocated = NULL;
2684                 btrfs_bio->end_io = NULL;
2685         }
2686         return bio;
2687 }
2688
2689 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2690 {
2691         struct btrfs_io_bio *btrfs_bio;
2692         struct bio *new;
2693
2694         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2695         if (new) {
2696                 btrfs_bio = btrfs_io_bio(new);
2697                 btrfs_bio->csum = NULL;
2698                 btrfs_bio->csum_allocated = NULL;
2699                 btrfs_bio->end_io = NULL;
2700         }
2701         return new;
2702 }
2703
2704 /* this also allocates from the btrfs_bioset */
2705 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2706 {
2707         struct btrfs_io_bio *btrfs_bio;
2708         struct bio *bio;
2709
2710         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2711         if (bio) {
2712                 btrfs_bio = btrfs_io_bio(bio);
2713                 btrfs_bio->csum = NULL;
2714                 btrfs_bio->csum_allocated = NULL;
2715                 btrfs_bio->end_io = NULL;
2716         }
2717         return bio;
2718 }
2719
2720
2721 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2722                                        unsigned long bio_flags)
2723 {
2724         int ret = 0;
2725         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2726         struct page *page = bvec->bv_page;
2727         struct extent_io_tree *tree = bio->bi_private;
2728         u64 start;
2729
2730         start = page_offset(page) + bvec->bv_offset;
2731
2732         bio->bi_private = NULL;
2733         bio_get(bio);
2734
2735         if (tree->ops && tree->ops->submit_bio_hook)
2736                 ret = tree->ops->submit_bio_hook(page->mapping->host, bio,
2737                                            mirror_num, bio_flags, start);
2738         else
2739                 btrfsic_submit_bio(bio);
2740
2741         bio_put(bio);
2742         return ret;
2743 }
2744
2745 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2746                      unsigned long offset, size_t size, struct bio *bio,
2747                      unsigned long bio_flags)
2748 {
2749         int ret = 0;
2750         if (tree->ops && tree->ops->merge_bio_hook)
2751                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2752                                                 bio_flags);
2753         return ret;
2754
2755 }
2756
2757 static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree,
2758                               struct writeback_control *wbc,
2759                               struct page *page, sector_t sector,
2760                               size_t size, unsigned long offset,
2761                               struct block_device *bdev,
2762                               struct bio **bio_ret,
2763                               unsigned long max_pages,
2764                               bio_end_io_t end_io_func,
2765                               int mirror_num,
2766                               unsigned long prev_bio_flags,
2767                               unsigned long bio_flags,
2768                               bool force_bio_submit)
2769 {
2770         int ret = 0;
2771         struct bio *bio;
2772         int contig = 0;
2773         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2774         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2775
2776         if (bio_ret && *bio_ret) {
2777                 bio = *bio_ret;
2778                 if (old_compressed)
2779                         contig = bio->bi_iter.bi_sector == sector;
2780                 else
2781                         contig = bio_end_sector(bio) == sector;
2782
2783                 if (prev_bio_flags != bio_flags || !contig ||
2784                     force_bio_submit ||
2785                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2786                     bio_add_page(bio, page, page_size, offset) < page_size) {
2787                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2788                         if (ret < 0) {
2789                                 *bio_ret = NULL;
2790                                 return ret;
2791                         }
2792                         bio = NULL;
2793                 } else {
2794                         if (wbc)
2795                                 wbc_account_io(wbc, page, page_size);
2796                         return 0;
2797                 }
2798         }
2799
2800         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2801                         GFP_NOFS | __GFP_HIGH);
2802         if (!bio)
2803                 return -ENOMEM;
2804
2805         bio_add_page(bio, page, page_size, offset);
2806         bio->bi_end_io = end_io_func;
2807         bio->bi_private = tree;
2808         bio_set_op_attrs(bio, op, op_flags);
2809         if (wbc) {
2810                 wbc_init_bio(wbc, bio);
2811                 wbc_account_io(wbc, page, page_size);
2812         }
2813
2814         if (bio_ret)
2815                 *bio_ret = bio;
2816         else
2817                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2818
2819         return ret;
2820 }
2821
2822 static void attach_extent_buffer_page(struct extent_buffer *eb,
2823                                       struct page *page)
2824 {
2825         if (!PagePrivate(page)) {
2826                 SetPagePrivate(page);
2827                 get_page(page);
2828                 set_page_private(page, (unsigned long)eb);
2829         } else {
2830                 WARN_ON(page->private != (unsigned long)eb);
2831         }
2832 }
2833
2834 void set_page_extent_mapped(struct page *page)
2835 {
2836         if (!PagePrivate(page)) {
2837                 SetPagePrivate(page);
2838                 get_page(page);
2839                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2840         }
2841 }
2842
2843 static struct extent_map *
2844 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2845                  u64 start, u64 len, get_extent_t *get_extent,
2846                  struct extent_map **em_cached)
2847 {
2848         struct extent_map *em;
2849
2850         if (em_cached && *em_cached) {
2851                 em = *em_cached;
2852                 if (extent_map_in_tree(em) && start >= em->start &&
2853                     start < extent_map_end(em)) {
2854                         atomic_inc(&em->refs);
2855                         return em;
2856                 }
2857
2858                 free_extent_map(em);
2859                 *em_cached = NULL;
2860         }
2861
2862         em = get_extent(inode, page, pg_offset, start, len, 0);
2863         if (em_cached && !IS_ERR_OR_NULL(em)) {
2864                 BUG_ON(*em_cached);
2865                 atomic_inc(&em->refs);
2866                 *em_cached = em;
2867         }
2868         return em;
2869 }
2870 /*
2871  * basic readpage implementation.  Locked extent state structs are inserted
2872  * into the tree that are removed when the IO is done (by the end_io
2873  * handlers)
2874  * XXX JDM: This needs looking at to ensure proper page locking
2875  * return 0 on success, otherwise return error
2876  */
2877 static int __do_readpage(struct extent_io_tree *tree,
2878                          struct page *page,
2879                          get_extent_t *get_extent,
2880                          struct extent_map **em_cached,
2881                          struct bio **bio, int mirror_num,
2882                          unsigned long *bio_flags, int read_flags,
2883                          u64 *prev_em_start)
2884 {
2885         struct inode *inode = page->mapping->host;
2886         u64 start = page_offset(page);
2887         u64 page_end = start + PAGE_SIZE - 1;
2888         u64 end;
2889         u64 cur = start;
2890         u64 extent_offset;
2891         u64 last_byte = i_size_read(inode);
2892         u64 block_start;
2893         u64 cur_end;
2894         sector_t sector;
2895         struct extent_map *em;
2896         struct block_device *bdev;
2897         int ret = 0;
2898         int nr = 0;
2899         size_t pg_offset = 0;
2900         size_t iosize;
2901         size_t disk_io_size;
2902         size_t blocksize = inode->i_sb->s_blocksize;
2903         unsigned long this_bio_flag = 0;
2904
2905         set_page_extent_mapped(page);
2906
2907         end = page_end;
2908         if (!PageUptodate(page)) {
2909                 if (cleancache_get_page(page) == 0) {
2910                         BUG_ON(blocksize != PAGE_SIZE);
2911                         unlock_extent(tree, start, end);
2912                         goto out;
2913                 }
2914         }
2915
2916         if (page->index == last_byte >> PAGE_SHIFT) {
2917                 char *userpage;
2918                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2919
2920                 if (zero_offset) {
2921                         iosize = PAGE_SIZE - zero_offset;
2922                         userpage = kmap_atomic(page);
2923                         memset(userpage + zero_offset, 0, iosize);
2924                         flush_dcache_page(page);
2925                         kunmap_atomic(userpage);
2926                 }
2927         }
2928         while (cur <= end) {
2929                 unsigned long pnr = (last_byte >> PAGE_SHIFT) + 1;
2930                 bool force_bio_submit = false;
2931
2932                 if (cur >= last_byte) {
2933                         char *userpage;
2934                         struct extent_state *cached = NULL;
2935
2936                         iosize = PAGE_SIZE - pg_offset;
2937                         userpage = kmap_atomic(page);
2938                         memset(userpage + pg_offset, 0, iosize);
2939                         flush_dcache_page(page);
2940                         kunmap_atomic(userpage);
2941                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2942                                             &cached, GFP_NOFS);
2943                         unlock_extent_cached(tree, cur,
2944                                              cur + iosize - 1,
2945                                              &cached, GFP_NOFS);
2946                         break;
2947                 }
2948                 em = __get_extent_map(inode, page, pg_offset, cur,
2949                                       end - cur + 1, get_extent, em_cached);
2950                 if (IS_ERR_OR_NULL(em)) {
2951                         SetPageError(page);
2952                         unlock_extent(tree, cur, end);
2953                         break;
2954                 }
2955                 extent_offset = cur - em->start;
2956                 BUG_ON(extent_map_end(em) <= cur);
2957                 BUG_ON(end < cur);
2958
2959                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2960                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2961                         extent_set_compress_type(&this_bio_flag,
2962                                                  em->compress_type);
2963                 }
2964
2965                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2966                 cur_end = min(extent_map_end(em) - 1, end);
2967                 iosize = ALIGN(iosize, blocksize);
2968                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2969                         disk_io_size = em->block_len;
2970                         sector = em->block_start >> 9;
2971                 } else {
2972                         sector = (em->block_start + extent_offset) >> 9;
2973                         disk_io_size = iosize;
2974                 }
2975                 bdev = em->bdev;
2976                 block_start = em->block_start;
2977                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2978                         block_start = EXTENT_MAP_HOLE;
2979
2980                 /*
2981                  * If we have a file range that points to a compressed extent
2982                  * and it's followed by a consecutive file range that points to
2983                  * to the same compressed extent (possibly with a different
2984                  * offset and/or length, so it either points to the whole extent
2985                  * or only part of it), we must make sure we do not submit a
2986                  * single bio to populate the pages for the 2 ranges because
2987                  * this makes the compressed extent read zero out the pages
2988                  * belonging to the 2nd range. Imagine the following scenario:
2989                  *
2990                  *  File layout
2991                  *  [0 - 8K]                     [8K - 24K]
2992                  *    |                               |
2993                  *    |                               |
2994                  * points to extent X,         points to extent X,
2995                  * offset 4K, length of 8K     offset 0, length 16K
2996                  *
2997                  * [extent X, compressed length = 4K uncompressed length = 16K]
2998                  *
2999                  * If the bio to read the compressed extent covers both ranges,
3000                  * it will decompress extent X into the pages belonging to the
3001                  * first range and then it will stop, zeroing out the remaining
3002                  * pages that belong to the other range that points to extent X.
3003                  * So here we make sure we submit 2 bios, one for the first
3004                  * range and another one for the third range. Both will target
3005                  * the same physical extent from disk, but we can't currently
3006                  * make the compressed bio endio callback populate the pages
3007                  * for both ranges because each compressed bio is tightly
3008                  * coupled with a single extent map, and each range can have
3009                  * an extent map with a different offset value relative to the
3010                  * uncompressed data of our extent and different lengths. This
3011                  * is a corner case so we prioritize correctness over
3012                  * non-optimal behavior (submitting 2 bios for the same extent).
3013                  */
3014                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3015                     prev_em_start && *prev_em_start != (u64)-1 &&
3016                     *prev_em_start != em->orig_start)
3017                         force_bio_submit = true;
3018
3019                 if (prev_em_start)
3020                         *prev_em_start = em->orig_start;
3021
3022                 free_extent_map(em);
3023                 em = NULL;
3024
3025                 /* we've found a hole, just zero and go on */
3026                 if (block_start == EXTENT_MAP_HOLE) {
3027                         char *userpage;
3028                         struct extent_state *cached = NULL;
3029
3030                         userpage = kmap_atomic(page);
3031                         memset(userpage + pg_offset, 0, iosize);
3032                         flush_dcache_page(page);
3033                         kunmap_atomic(userpage);
3034
3035                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3036                                             &cached, GFP_NOFS);
3037                         unlock_extent_cached(tree, cur,
3038                                              cur + iosize - 1,
3039                                              &cached, GFP_NOFS);
3040                         cur = cur + iosize;
3041                         pg_offset += iosize;
3042                         continue;
3043                 }
3044                 /* the get_extent function already copied into the page */
3045                 if (test_range_bit(tree, cur, cur_end,
3046                                    EXTENT_UPTODATE, 1, NULL)) {
3047                         check_page_uptodate(tree, page);
3048                         unlock_extent(tree, cur, cur + iosize - 1);
3049                         cur = cur + iosize;
3050                         pg_offset += iosize;
3051                         continue;
3052                 }
3053                 /* we have an inline extent but it didn't get marked up
3054                  * to date.  Error out
3055                  */
3056                 if (block_start == EXTENT_MAP_INLINE) {
3057                         SetPageError(page);
3058                         unlock_extent(tree, cur, cur + iosize - 1);
3059                         cur = cur + iosize;
3060                         pg_offset += iosize;
3061                         continue;
3062                 }
3063
3064                 pnr -= page->index;
3065                 ret = submit_extent_page(REQ_OP_READ, read_flags, tree, NULL,
3066                                          page, sector, disk_io_size, pg_offset,
3067                                          bdev, bio, pnr,
3068                                          end_bio_extent_readpage, mirror_num,
3069                                          *bio_flags,
3070                                          this_bio_flag,
3071                                          force_bio_submit);
3072                 if (!ret) {
3073                         nr++;
3074                         *bio_flags = this_bio_flag;
3075                 } else {
3076                         SetPageError(page);
3077                         unlock_extent(tree, cur, cur + iosize - 1);
3078                         goto out;
3079                 }
3080                 cur = cur + iosize;
3081                 pg_offset += iosize;
3082         }
3083 out:
3084         if (!nr) {
3085                 if (!PageError(page))
3086                         SetPageUptodate(page);
3087                 unlock_page(page);
3088         }
3089         return ret;
3090 }
3091
3092 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3093                                              struct page *pages[], int nr_pages,
3094                                              u64 start, u64 end,
3095                                              get_extent_t *get_extent,
3096                                              struct extent_map **em_cached,
3097                                              struct bio **bio, int mirror_num,
3098                                              unsigned long *bio_flags,
3099                                              u64 *prev_em_start)
3100 {
3101         struct inode *inode;
3102         struct btrfs_ordered_extent *ordered;
3103         int index;
3104
3105         inode = pages[0]->mapping->host;
3106         while (1) {
3107                 lock_extent(tree, start, end);
3108                 ordered = btrfs_lookup_ordered_range(inode, start,
3109                                                      end - start + 1);
3110                 if (!ordered)
3111                         break;
3112                 unlock_extent(tree, start, end);
3113                 btrfs_start_ordered_extent(inode, ordered, 1);
3114                 btrfs_put_ordered_extent(ordered);
3115         }
3116
3117         for (index = 0; index < nr_pages; index++) {
3118                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3119                               mirror_num, bio_flags, 0, prev_em_start);
3120                 put_page(pages[index]);
3121         }
3122 }
3123
3124 static void __extent_readpages(struct extent_io_tree *tree,
3125                                struct page *pages[],
3126                                int nr_pages, get_extent_t *get_extent,
3127                                struct extent_map **em_cached,
3128                                struct bio **bio, int mirror_num,
3129                                unsigned long *bio_flags,
3130                                u64 *prev_em_start)
3131 {
3132         u64 start = 0;
3133         u64 end = 0;
3134         u64 page_start;
3135         int index;
3136         int first_index = 0;
3137
3138         for (index = 0; index < nr_pages; index++) {
3139                 page_start = page_offset(pages[index]);
3140                 if (!end) {
3141                         start = page_start;
3142                         end = start + PAGE_SIZE - 1;
3143                         first_index = index;
3144                 } else if (end + 1 == page_start) {
3145                         end += PAGE_SIZE;
3146                 } else {
3147                         __do_contiguous_readpages(tree, &pages[first_index],
3148                                                   index - first_index, start,
3149                                                   end, get_extent, em_cached,
3150                                                   bio, mirror_num, bio_flags,
3151                                                   prev_em_start);
3152                         start = page_start;
3153                         end = start + PAGE_SIZE - 1;
3154                         first_index = index;
3155                 }
3156         }
3157
3158         if (end)
3159                 __do_contiguous_readpages(tree, &pages[first_index],
3160                                           index - first_index, start,
3161                                           end, get_extent, em_cached, bio,
3162                                           mirror_num, bio_flags,
3163                                           prev_em_start);
3164 }
3165
3166 static int __extent_read_full_page(struct extent_io_tree *tree,
3167                                    struct page *page,
3168                                    get_extent_t *get_extent,
3169                                    struct bio **bio, int mirror_num,
3170                                    unsigned long *bio_flags, int read_flags)
3171 {
3172         struct inode *inode = page->mapping->host;
3173         struct btrfs_ordered_extent *ordered;
3174         u64 start = page_offset(page);
3175         u64 end = start + PAGE_SIZE - 1;
3176         int ret;
3177
3178         while (1) {
3179                 lock_extent(tree, start, end);
3180                 ordered = btrfs_lookup_ordered_range(inode, start,
3181                                                 PAGE_SIZE);
3182                 if (!ordered)
3183                         break;
3184                 unlock_extent(tree, start, end);
3185                 btrfs_start_ordered_extent(inode, ordered, 1);
3186                 btrfs_put_ordered_extent(ordered);
3187         }
3188
3189         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3190                             bio_flags, read_flags, NULL);
3191         return ret;
3192 }
3193
3194 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3195                             get_extent_t *get_extent, int mirror_num)
3196 {
3197         struct bio *bio = NULL;
3198         unsigned long bio_flags = 0;
3199         int ret;
3200
3201         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3202                                       &bio_flags, 0);
3203         if (bio)
3204                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3205         return ret;
3206 }
3207
3208 static void update_nr_written(struct page *page, struct writeback_control *wbc,
3209                               unsigned long nr_written)
3210 {
3211         wbc->nr_to_write -= nr_written;
3212 }
3213
3214 /*
3215  * helper for __extent_writepage, doing all of the delayed allocation setup.
3216  *
3217  * This returns 1 if our fill_delalloc function did all the work required
3218  * to write the page (copy into inline extent).  In this case the IO has
3219  * been started and the page is already unlocked.
3220  *
3221  * This returns 0 if all went well (page still locked)
3222  * This returns < 0 if there were errors (page still locked)
3223  */
3224 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3225                               struct page *page, struct writeback_control *wbc,
3226                               struct extent_page_data *epd,
3227                               u64 delalloc_start,
3228                               unsigned long *nr_written)
3229 {
3230         struct extent_io_tree *tree = epd->tree;
3231         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3232         u64 nr_delalloc;
3233         u64 delalloc_to_write = 0;
3234         u64 delalloc_end = 0;
3235         int ret;
3236         int page_started = 0;
3237
3238         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3239                 return 0;
3240
3241         while (delalloc_end < page_end) {
3242                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3243                                                page,
3244                                                &delalloc_start,
3245                                                &delalloc_end,
3246                                                BTRFS_MAX_EXTENT_SIZE);
3247                 if (nr_delalloc == 0) {
3248                         delalloc_start = delalloc_end + 1;
3249                         continue;
3250                 }
3251                 ret = tree->ops->fill_delalloc(inode, page,
3252                                                delalloc_start,
3253                                                delalloc_end,
3254                                                &page_started,
3255                                                nr_written);
3256                 /* File system has been set read-only */
3257                 if (ret) {
3258                         SetPageError(page);
3259                         /* fill_delalloc should be return < 0 for error
3260                          * but just in case, we use > 0 here meaning the
3261                          * IO is started, so we don't want to return > 0
3262                          * unless things are going well.
3263                          */
3264                         ret = ret < 0 ? ret : -EIO;
3265                         goto done;
3266                 }
3267                 /*
3268                  * delalloc_end is already one less than the total length, so
3269                  * we don't subtract one from PAGE_SIZE
3270                  */
3271                 delalloc_to_write += (delalloc_end - delalloc_start +
3272                                       PAGE_SIZE) >> PAGE_SHIFT;
3273                 delalloc_start = delalloc_end + 1;
3274         }
3275         if (wbc->nr_to_write < delalloc_to_write) {
3276                 int thresh = 8192;
3277
3278                 if (delalloc_to_write < thresh * 2)
3279                         thresh = delalloc_to_write;
3280                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3281                                          thresh);
3282         }
3283
3284         /* did the fill delalloc function already unlock and start
3285          * the IO?
3286          */
3287         if (page_started) {
3288                 /*
3289                  * we've unlocked the page, so we can't update
3290                  * the mapping's writeback index, just update
3291                  * nr_to_write.
3292                  */
3293                 wbc->nr_to_write -= *nr_written;
3294                 return 1;
3295         }
3296
3297         ret = 0;
3298
3299 done:
3300         return ret;
3301 }
3302
3303 /*
3304  * helper for __extent_writepage.  This calls the writepage start hooks,
3305  * and does the loop to map the page into extents and bios.
3306  *
3307  * We return 1 if the IO is started and the page is unlocked,
3308  * 0 if all went well (page still locked)
3309  * < 0 if there were errors (page still locked)
3310  */
3311 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3312                                  struct page *page,
3313                                  struct writeback_control *wbc,
3314                                  struct extent_page_data *epd,
3315                                  loff_t i_size,
3316                                  unsigned long nr_written,
3317                                  int write_flags, int *nr_ret)
3318 {
3319         struct extent_io_tree *tree = epd->tree;
3320         u64 start = page_offset(page);
3321         u64 page_end = start + PAGE_SIZE - 1;
3322         u64 end;
3323         u64 cur = start;
3324         u64 extent_offset;
3325         u64 block_start;
3326         u64 iosize;
3327         sector_t sector;
3328         struct extent_state *cached_state = NULL;
3329         struct extent_map *em;
3330         struct block_device *bdev;
3331         size_t pg_offset = 0;
3332         size_t blocksize;
3333         int ret = 0;
3334         int nr = 0;
3335         bool compressed;
3336
3337         if (tree->ops && tree->ops->writepage_start_hook) {
3338                 ret = tree->ops->writepage_start_hook(page, start,
3339                                                       page_end);
3340                 if (ret) {
3341                         /* Fixup worker will requeue */
3342                         if (ret == -EBUSY)
3343                                 wbc->pages_skipped++;
3344                         else
3345                                 redirty_page_for_writepage(wbc, page);
3346
3347                         update_nr_written(page, wbc, nr_written);
3348                         unlock_page(page);
3349                         ret = 1;
3350                         goto done_unlocked;
3351                 }
3352         }
3353
3354         /*
3355          * we don't want to touch the inode after unlocking the page,
3356          * so we update the mapping writeback index now
3357          */
3358         update_nr_written(page, wbc, nr_written + 1);
3359
3360         end = page_end;
3361         if (i_size <= start) {
3362                 if (tree->ops && tree->ops->writepage_end_io_hook)
3363                         tree->ops->writepage_end_io_hook(page, start,
3364                                                          page_end, NULL, 1);
3365                 goto done;
3366         }
3367
3368         blocksize = inode->i_sb->s_blocksize;
3369
3370         while (cur <= end) {
3371                 u64 em_end;
3372                 unsigned long max_nr;
3373
3374                 if (cur >= i_size) {
3375                         if (tree->ops && tree->ops->writepage_end_io_hook)
3376                                 tree->ops->writepage_end_io_hook(page, cur,
3377                                                          page_end, NULL, 1);
3378                         break;
3379                 }
3380                 em = epd->get_extent(inode, page, pg_offset, cur,
3381                                      end - cur + 1, 1);
3382                 if (IS_ERR_OR_NULL(em)) {
3383                         SetPageError(page);
3384                         ret = PTR_ERR_OR_ZERO(em);
3385                         break;
3386                 }
3387
3388                 extent_offset = cur - em->start;
3389                 em_end = extent_map_end(em);
3390                 BUG_ON(em_end <= cur);
3391                 BUG_ON(end < cur);
3392                 iosize = min(em_end - cur, end - cur + 1);
3393                 iosize = ALIGN(iosize, blocksize);
3394                 sector = (em->block_start + extent_offset) >> 9;
3395                 bdev = em->bdev;
3396                 block_start = em->block_start;
3397                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3398                 free_extent_map(em);
3399                 em = NULL;
3400
3401                 /*
3402                  * compressed and inline extents are written through other
3403                  * paths in the FS
3404                  */
3405                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3406                     block_start == EXTENT_MAP_INLINE) {
3407                         /*
3408                          * end_io notification does not happen here for
3409                          * compressed extents
3410                          */
3411                         if (!compressed && tree->ops &&
3412                             tree->ops->writepage_end_io_hook)
3413                                 tree->ops->writepage_end_io_hook(page, cur,
3414                                                          cur + iosize - 1,
3415                                                          NULL, 1);
3416                         else if (compressed) {
3417                                 /* we don't want to end_page_writeback on
3418                                  * a compressed extent.  this happens
3419                                  * elsewhere
3420                                  */
3421                                 nr++;
3422                         }
3423
3424                         cur += iosize;
3425                         pg_offset += iosize;
3426                         continue;
3427                 }
3428
3429                 max_nr = (i_size >> PAGE_SHIFT) + 1;
3430
3431                 set_range_writeback(tree, cur, cur + iosize - 1);
3432                 if (!PageWriteback(page)) {
3433                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3434                                    "page %lu not writeback, cur %llu end %llu",
3435                                page->index, cur, end);
3436                 }
3437
3438                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3439                                          page, sector, iosize, pg_offset,
3440                                          bdev, &epd->bio, max_nr,
3441                                          end_bio_extent_writepage,
3442                                          0, 0, 0, false);
3443                 if (ret)
3444                         SetPageError(page);
3445
3446                 cur = cur + iosize;
3447                 pg_offset += iosize;
3448                 nr++;
3449         }
3450 done:
3451         *nr_ret = nr;
3452
3453 done_unlocked:
3454
3455         /* drop our reference on any cached states */
3456         free_extent_state(cached_state);
3457         return ret;
3458 }
3459
3460 /*
3461  * the writepage semantics are similar to regular writepage.  extent
3462  * records are inserted to lock ranges in the tree, and as dirty areas
3463  * are found, they are marked writeback.  Then the lock bits are removed
3464  * and the end_io handler clears the writeback ranges
3465  */
3466 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3467                               void *data)
3468 {
3469         struct inode *inode = page->mapping->host;
3470         struct extent_page_data *epd = data;
3471         u64 start = page_offset(page);
3472         u64 page_end = start + PAGE_SIZE - 1;
3473         int ret;
3474         int nr = 0;
3475         size_t pg_offset = 0;
3476         loff_t i_size = i_size_read(inode);
3477         unsigned long end_index = i_size >> PAGE_SHIFT;
3478         int write_flags = 0;
3479         unsigned long nr_written = 0;
3480
3481         if (wbc->sync_mode == WB_SYNC_ALL)
3482                 write_flags = WRITE_SYNC;
3483
3484         trace___extent_writepage(page, inode, wbc);
3485
3486         WARN_ON(!PageLocked(page));
3487
3488         ClearPageError(page);
3489
3490         pg_offset = i_size & (PAGE_SIZE - 1);
3491         if (page->index > end_index ||
3492            (page->index == end_index && !pg_offset)) {
3493                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3494                 unlock_page(page);
3495                 return 0;
3496         }
3497
3498         if (page->index == end_index) {
3499                 char *userpage;
3500
3501                 userpage = kmap_atomic(page);
3502                 memset(userpage + pg_offset, 0,
3503                        PAGE_SIZE - pg_offset);
3504                 kunmap_atomic(userpage);
3505                 flush_dcache_page(page);
3506         }
3507
3508         pg_offset = 0;
3509
3510         set_page_extent_mapped(page);
3511
3512         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3513         if (ret == 1)
3514                 goto done_unlocked;
3515         if (ret)
3516                 goto done;
3517
3518         ret = __extent_writepage_io(inode, page, wbc, epd,
3519                                     i_size, nr_written, write_flags, &nr);
3520         if (ret == 1)
3521                 goto done_unlocked;
3522
3523 done:
3524         if (nr == 0) {
3525                 /* make sure the mapping tag for page dirty gets cleared */
3526                 set_page_writeback(page);
3527                 end_page_writeback(page);
3528         }
3529         if (PageError(page)) {
3530                 ret = ret < 0 ? ret : -EIO;
3531                 end_extent_writepage(page, ret, start, page_end);
3532         }
3533         unlock_page(page);
3534         return ret;
3535
3536 done_unlocked:
3537         return 0;
3538 }
3539
3540 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3541 {
3542         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3543                        TASK_UNINTERRUPTIBLE);
3544 }
3545
3546 static noinline_for_stack int
3547 lock_extent_buffer_for_io(struct extent_buffer *eb,
3548                           struct btrfs_fs_info *fs_info,
3549                           struct extent_page_data *epd)
3550 {
3551         unsigned long i, num_pages;
3552         int flush = 0;
3553         int ret = 0;
3554
3555         if (!btrfs_try_tree_write_lock(eb)) {
3556                 flush = 1;
3557                 flush_write_bio(epd);
3558                 btrfs_tree_lock(eb);
3559         }
3560
3561         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3562                 btrfs_tree_unlock(eb);
3563                 if (!epd->sync_io)
3564                         return 0;
3565                 if (!flush) {
3566                         flush_write_bio(epd);
3567                         flush = 1;
3568                 }
3569                 while (1) {
3570                         wait_on_extent_buffer_writeback(eb);
3571                         btrfs_tree_lock(eb);
3572                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3573                                 break;
3574                         btrfs_tree_unlock(eb);
3575                 }
3576         }
3577
3578         /*
3579          * We need to do this to prevent races in people who check if the eb is
3580          * under IO since we can end up having no IO bits set for a short period
3581          * of time.
3582          */
3583         spin_lock(&eb->refs_lock);
3584         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3585                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3586                 spin_unlock(&eb->refs_lock);
3587                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3588                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3589                                      -eb->len,
3590                                      fs_info->dirty_metadata_batch);
3591                 ret = 1;
3592         } else {
3593                 spin_unlock(&eb->refs_lock);
3594         }
3595
3596         btrfs_tree_unlock(eb);
3597
3598         if (!ret)
3599                 return ret;
3600
3601         num_pages = num_extent_pages(eb->start, eb->len);
3602         for (i = 0; i < num_pages; i++) {
3603                 struct page *p = eb->pages[i];
3604
3605                 if (!trylock_page(p)) {
3606                         if (!flush) {
3607                                 flush_write_bio(epd);
3608                                 flush = 1;
3609                         }
3610                         lock_page(p);
3611                 }
3612         }
3613
3614         return ret;
3615 }
3616
3617 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3618 {
3619         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3620         smp_mb__after_atomic();
3621         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3622 }
3623
3624 static void set_btree_ioerr(struct page *page)
3625 {
3626         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3627         struct btrfs_inode *btree_ino = BTRFS_I(eb->fs_info->btree_inode);
3628
3629         SetPageError(page);
3630         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3631                 return;
3632
3633         /*
3634          * If writeback for a btree extent that doesn't belong to a log tree
3635          * failed, increment the counter transaction->eb_write_errors.
3636          * We do this because while the transaction is running and before it's
3637          * committing (when we call filemap_fdata[write|wait]_range against
3638          * the btree inode), we might have
3639          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3640          * returns an error or an error happens during writeback, when we're
3641          * committing the transaction we wouldn't know about it, since the pages
3642          * can be no longer dirty nor marked anymore for writeback (if a
3643          * subsequent modification to the extent buffer didn't happen before the
3644          * transaction commit), which makes filemap_fdata[write|wait]_range not
3645          * able to find the pages tagged with SetPageError at transaction
3646          * commit time. So if this happens we must abort the transaction,
3647          * otherwise we commit a super block with btree roots that point to
3648          * btree nodes/leafs whose content on disk is invalid - either garbage
3649          * or the content of some node/leaf from a past generation that got
3650          * cowed or deleted and is no longer valid.
3651          *
3652          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3653          * not be enough - we need to distinguish between log tree extents vs
3654          * non-log tree extents, and the next filemap_fdatawait_range() call
3655          * will catch and clear such errors in the mapping - and that call might
3656          * be from a log sync and not from a transaction commit. Also, checking
3657          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3658          * not done and would not be reliable - the eb might have been released
3659          * from memory and reading it back again means that flag would not be
3660          * set (since it's a runtime flag, not persisted on disk).
3661          *
3662          * Using the flags below in the btree inode also makes us achieve the
3663          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3664          * writeback for all dirty pages and before filemap_fdatawait_range()
3665          * is called, the writeback for all dirty pages had already finished
3666          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3667          * filemap_fdatawait_range() would return success, as it could not know
3668          * that writeback errors happened (the pages were no longer tagged for
3669          * writeback).
3670          */
3671         switch (eb->log_index) {
3672         case -1:
3673                 set_bit(BTRFS_INODE_BTREE_ERR, &btree_ino->runtime_flags);
3674                 break;
3675         case 0:
3676                 set_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
3677                 break;
3678         case 1:
3679                 set_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
3680                 break;
3681         default:
3682                 BUG(); /* unexpected, logic error */
3683         }
3684 }
3685
3686 static void end_bio_extent_buffer_writepage(struct bio *bio)
3687 {
3688         struct bio_vec *bvec;
3689         struct extent_buffer *eb;
3690         int i, done;
3691
3692         bio_for_each_segment_all(bvec, bio, i) {
3693                 struct page *page = bvec->bv_page;
3694
3695                 eb = (struct extent_buffer *)page->private;
3696                 BUG_ON(!eb);
3697                 done = atomic_dec_and_test(&eb->io_pages);
3698
3699                 if (bio->bi_error ||
3700                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3701                         ClearPageUptodate(page);
3702                         set_btree_ioerr(page);
3703                 }
3704
3705                 end_page_writeback(page);
3706
3707                 if (!done)
3708                         continue;
3709
3710                 end_extent_buffer_writeback(eb);
3711         }
3712
3713         bio_put(bio);
3714 }
3715
3716 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3717                         struct btrfs_fs_info *fs_info,
3718                         struct writeback_control *wbc,
3719                         struct extent_page_data *epd)
3720 {
3721         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3722         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3723         u64 offset = eb->start;
3724         unsigned long i, num_pages;
3725         unsigned long bio_flags = 0;
3726         int write_flags = (epd->sync_io ? WRITE_SYNC : 0) | REQ_META;
3727         int ret = 0;
3728
3729         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3730         num_pages = num_extent_pages(eb->start, eb->len);
3731         atomic_set(&eb->io_pages, num_pages);
3732         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3733                 bio_flags = EXTENT_BIO_TREE_LOG;
3734
3735         for (i = 0; i < num_pages; i++) {
3736                 struct page *p = eb->pages[i];
3737
3738                 clear_page_dirty_for_io(p);
3739                 set_page_writeback(p);
3740                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3741                                          p, offset >> 9, PAGE_SIZE, 0, bdev,
3742                                          &epd->bio, -1,
3743                                          end_bio_extent_buffer_writepage,
3744                                          0, epd->bio_flags, bio_flags, false);
3745                 epd->bio_flags = bio_flags;
3746                 if (ret) {
3747                         set_btree_ioerr(p);
3748                         end_page_writeback(p);
3749                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3750                                 end_extent_buffer_writeback(eb);
3751                         ret = -EIO;
3752                         break;
3753                 }
3754                 offset += PAGE_SIZE;
3755                 update_nr_written(p, wbc, 1);
3756                 unlock_page(p);
3757         }
3758
3759         if (unlikely(ret)) {
3760                 for (; i < num_pages; i++) {
3761                         struct page *p = eb->pages[i];
3762                         clear_page_dirty_for_io(p);
3763                         unlock_page(p);
3764                 }
3765         }
3766
3767         return ret;
3768 }
3769
3770 int btree_write_cache_pages(struct address_space *mapping,
3771                                    struct writeback_control *wbc)
3772 {
3773         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3774         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3775         struct extent_buffer *eb, *prev_eb = NULL;
3776         struct extent_page_data epd = {
3777                 .bio = NULL,
3778                 .tree = tree,
3779                 .extent_locked = 0,
3780                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3781                 .bio_flags = 0,
3782         };
3783         int ret = 0;
3784         int done = 0;
3785         int nr_to_write_done = 0;
3786         struct pagevec pvec;
3787         int nr_pages;
3788         pgoff_t index;
3789         pgoff_t end;            /* Inclusive */
3790         int scanned = 0;
3791         int tag;
3792
3793         pagevec_init(&pvec, 0);
3794         if (wbc->range_cyclic) {
3795                 index = mapping->writeback_index; /* Start from prev offset */
3796                 end = -1;
3797         } else {
3798                 index = wbc->range_start >> PAGE_SHIFT;
3799                 end = wbc->range_end >> PAGE_SHIFT;
3800                 scanned = 1;
3801         }
3802         if (wbc->sync_mode == WB_SYNC_ALL)
3803                 tag = PAGECACHE_TAG_TOWRITE;
3804         else
3805                 tag = PAGECACHE_TAG_DIRTY;
3806 retry:
3807         if (wbc->sync_mode == WB_SYNC_ALL)
3808                 tag_pages_for_writeback(mapping, index, end);
3809         while (!done && !nr_to_write_done && (index <= end) &&
3810                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3811                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3812                 unsigned i;
3813
3814                 scanned = 1;
3815                 for (i = 0; i < nr_pages; i++) {
3816                         struct page *page = pvec.pages[i];
3817
3818                         if (!PagePrivate(page))
3819                                 continue;
3820
3821                         if (!wbc->range_cyclic && page->index > end) {
3822                                 done = 1;
3823                                 break;
3824                         }
3825
3826                         spin_lock(&mapping->private_lock);
3827                         if (!PagePrivate(page)) {
3828                                 spin_unlock(&mapping->private_lock);
3829                                 continue;
3830                         }
3831
3832                         eb = (struct extent_buffer *)page->private;
3833
3834                         /*
3835                          * Shouldn't happen and normally this would be a BUG_ON
3836                          * but no sense in crashing the users box for something
3837                          * we can survive anyway.
3838                          */
3839                         if (WARN_ON(!eb)) {
3840                                 spin_unlock(&mapping->private_lock);
3841                                 continue;
3842                         }
3843
3844                         if (eb == prev_eb) {
3845                                 spin_unlock(&mapping->private_lock);
3846                                 continue;
3847                         }
3848
3849                         ret = atomic_inc_not_zero(&eb->refs);
3850                         spin_unlock(&mapping->private_lock);
3851                         if (!ret)
3852                                 continue;
3853
3854                         prev_eb = eb;
3855                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3856                         if (!ret) {
3857                                 free_extent_buffer(eb);
3858                                 continue;
3859                         }
3860
3861                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3862                         if (ret) {
3863                                 done = 1;
3864                                 free_extent_buffer(eb);
3865                                 break;
3866                         }
3867                         free_extent_buffer(eb);
3868
3869                         /*
3870                          * the filesystem may choose to bump up nr_to_write.
3871                          * We have to make sure to honor the new nr_to_write
3872                          * at any time
3873                          */
3874                         nr_to_write_done = wbc->nr_to_write <= 0;
3875                 }
3876                 pagevec_release(&pvec);
3877                 cond_resched();
3878         }
3879         if (!scanned && !done) {
3880                 /*
3881                  * We hit the last page and there is more work to be done: wrap
3882                  * back to the start of the file
3883                  */
3884                 scanned = 1;
3885                 index = 0;
3886                 goto retry;
3887         }
3888         flush_write_bio(&epd);
3889         return ret;
3890 }
3891
3892 /**
3893  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3894  * @mapping: address space structure to write
3895  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3896  * @writepage: function called for each page
3897  * @data: data passed to writepage function
3898  *
3899  * If a page is already under I/O, write_cache_pages() skips it, even
3900  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3901  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3902  * and msync() need to guarantee that all the data which was dirty at the time
3903  * the call was made get new I/O started against them.  If wbc->sync_mode is
3904  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3905  * existing IO to complete.
3906  */
3907 static int extent_write_cache_pages(struct extent_io_tree *tree,
3908                              struct address_space *mapping,
3909                              struct writeback_control *wbc,
3910                              writepage_t writepage, void *data,
3911                              void (*flush_fn)(void *))
3912 {
3913         struct inode *inode = mapping->host;
3914         int ret = 0;
3915         int done = 0;
3916         int nr_to_write_done = 0;
3917         struct pagevec pvec;
3918         int nr_pages;
3919         pgoff_t index;
3920         pgoff_t end;            /* Inclusive */
3921         pgoff_t done_index;
3922         int range_whole = 0;
3923         int scanned = 0;
3924         int tag;
3925
3926         /*
3927          * We have to hold onto the inode so that ordered extents can do their
3928          * work when the IO finishes.  The alternative to this is failing to add
3929          * an ordered extent if the igrab() fails there and that is a huge pain
3930          * to deal with, so instead just hold onto the inode throughout the
3931          * writepages operation.  If it fails here we are freeing up the inode
3932          * anyway and we'd rather not waste our time writing out stuff that is
3933          * going to be truncated anyway.
3934          */
3935         if (!igrab(inode))
3936                 return 0;
3937
3938         pagevec_init(&pvec, 0);
3939         if (wbc->range_cyclic) {
3940                 index = mapping->writeback_index; /* Start from prev offset */
3941                 end = -1;
3942         } else {
3943                 index = wbc->range_start >> PAGE_SHIFT;
3944                 end = wbc->range_end >> PAGE_SHIFT;
3945                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3946                         range_whole = 1;
3947                 scanned = 1;
3948         }
3949         if (wbc->sync_mode == WB_SYNC_ALL)
3950                 tag = PAGECACHE_TAG_TOWRITE;
3951         else
3952                 tag = PAGECACHE_TAG_DIRTY;
3953 retry:
3954         if (wbc->sync_mode == WB_SYNC_ALL)
3955                 tag_pages_for_writeback(mapping, index, end);
3956         done_index = index;
3957         while (!done && !nr_to_write_done && (index <= end) &&
3958                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3959                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3960                 unsigned i;
3961
3962                 scanned = 1;
3963                 for (i = 0; i < nr_pages; i++) {
3964                         struct page *page = pvec.pages[i];
3965
3966                         done_index = page->index;
3967                         /*
3968                          * At this point we hold neither mapping->tree_lock nor
3969                          * lock on the page itself: the page may be truncated or
3970                          * invalidated (changing page->mapping to NULL), or even
3971                          * swizzled back from swapper_space to tmpfs file
3972                          * mapping
3973                          */
3974                         if (!trylock_page(page)) {
3975                                 flush_fn(data);
3976                                 lock_page(page);
3977                         }
3978
3979                         if (unlikely(page->mapping != mapping)) {
3980                                 unlock_page(page);
3981                                 continue;
3982                         }
3983
3984                         if (!wbc->range_cyclic && page->index > end) {
3985                                 done = 1;
3986                                 unlock_page(page);
3987                                 continue;
3988                         }
3989
3990                         if (wbc->sync_mode != WB_SYNC_NONE) {
3991                                 if (PageWriteback(page))
3992                                         flush_fn(data);
3993                                 wait_on_page_writeback(page);
3994                         }
3995
3996                         if (PageWriteback(page) ||
3997                             !clear_page_dirty_for_io(page)) {
3998                                 unlock_page(page);
3999                                 continue;
4000                         }
4001
4002                         ret = (*writepage)(page, wbc, data);
4003
4004                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4005                                 unlock_page(page);
4006                                 ret = 0;
4007                         }
4008                         if (ret < 0) {
4009                                 /*
4010                                  * done_index is set past this page,
4011                                  * so media errors will not choke
4012                                  * background writeout for the entire
4013                                  * file. This has consequences for
4014                                  * range_cyclic semantics (ie. it may
4015                                  * not be suitable for data integrity
4016                                  * writeout).
4017                                  */
4018                                 done_index = page->index + 1;
4019                                 done = 1;
4020                                 break;
4021                         }
4022
4023                         /*
4024                          * the filesystem may choose to bump up nr_to_write.
4025                          * We have to make sure to honor the new nr_to_write
4026                          * at any time
4027                          */
4028                         nr_to_write_done = wbc->nr_to_write <= 0;
4029                 }
4030                 pagevec_release(&pvec);
4031                 cond_resched();
4032         }
4033         if (!scanned && !done) {
4034                 /*
4035                  * We hit the last page and there is more work to be done: wrap
4036                  * back to the start of the file
4037                  */
4038                 scanned = 1;
4039                 index = 0;
4040                 goto retry;
4041         }
4042
4043         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4044                 mapping->writeback_index = done_index;
4045
4046         btrfs_add_delayed_iput(inode);
4047         return ret;
4048 }
4049
4050 static void flush_epd_write_bio(struct extent_page_data *epd)
4051 {
4052         if (epd->bio) {
4053                 int ret;
4054
4055                 bio_set_op_attrs(epd->bio, REQ_OP_WRITE,
4056                                  epd->sync_io ? WRITE_SYNC : 0);
4057
4058                 ret = submit_one_bio(epd->bio, 0, epd->bio_flags);
4059                 BUG_ON(ret < 0); /* -ENOMEM */
4060                 epd->bio = NULL;
4061         }
4062 }
4063
4064 static noinline void flush_write_bio(void *data)
4065 {
4066         struct extent_page_data *epd = data;
4067         flush_epd_write_bio(epd);
4068 }
4069
4070 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4071                           get_extent_t *get_extent,
4072                           struct writeback_control *wbc)
4073 {
4074         int ret;
4075         struct extent_page_data epd = {
4076                 .bio = NULL,
4077                 .tree = tree,
4078                 .get_extent = get_extent,
4079                 .extent_locked = 0,
4080                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4081                 .bio_flags = 0,
4082         };
4083
4084         ret = __extent_writepage(page, wbc, &epd);
4085
4086         flush_epd_write_bio(&epd);
4087         return ret;
4088 }
4089
4090 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4091                               u64 start, u64 end, get_extent_t *get_extent,
4092                               int mode)
4093 {
4094         int ret = 0;
4095         struct address_space *mapping = inode->i_mapping;
4096         struct page *page;
4097         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4098                 PAGE_SHIFT;
4099
4100         struct extent_page_data epd = {
4101                 .bio = NULL,
4102                 .tree = tree,
4103                 .get_extent = get_extent,
4104                 .extent_locked = 1,
4105                 .sync_io = mode == WB_SYNC_ALL,
4106                 .bio_flags = 0,
4107         };
4108         struct writeback_control wbc_writepages = {
4109                 .sync_mode      = mode,
4110                 .nr_to_write    = nr_pages * 2,
4111                 .range_start    = start,
4112                 .range_end      = end + 1,
4113         };
4114
4115         while (start <= end) {
4116                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4117                 if (clear_page_dirty_for_io(page))
4118                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4119                 else {
4120                         if (tree->ops && tree->ops->writepage_end_io_hook)
4121                                 tree->ops->writepage_end_io_hook(page, start,
4122                                                  start + PAGE_SIZE - 1,
4123                                                  NULL, 1);
4124                         unlock_page(page);
4125                 }
4126                 put_page(page);
4127                 start += PAGE_SIZE;
4128         }
4129
4130         flush_epd_write_bio(&epd);
4131         return ret;
4132 }
4133
4134 int extent_writepages(struct extent_io_tree *tree,
4135                       struct address_space *mapping,
4136                       get_extent_t *get_extent,
4137                       struct writeback_control *wbc)
4138 {
4139         int ret = 0;
4140         struct extent_page_data epd = {
4141                 .bio = NULL,
4142                 .tree = tree,
4143                 .get_extent = get_extent,
4144                 .extent_locked = 0,
4145                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4146                 .bio_flags = 0,
4147         };
4148
4149         ret = extent_write_cache_pages(tree, mapping, wbc,
4150                                        __extent_writepage, &epd,
4151                                        flush_write_bio);
4152         flush_epd_write_bio(&epd);
4153         return ret;
4154 }
4155
4156 int extent_readpages(struct extent_io_tree *tree,
4157                      struct address_space *mapping,
4158                      struct list_head *pages, unsigned nr_pages,
4159                      get_extent_t get_extent)
4160 {
4161         struct bio *bio = NULL;
4162         unsigned page_idx;
4163         unsigned long bio_flags = 0;
4164         struct page *pagepool[16];
4165         struct page *page;
4166         struct extent_map *em_cached = NULL;
4167         int nr = 0;
4168         u64 prev_em_start = (u64)-1;
4169
4170         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4171                 page = list_entry(pages->prev, struct page, lru);
4172
4173                 prefetchw(&page->flags);
4174                 list_del(&page->lru);
4175                 if (add_to_page_cache_lru(page, mapping,
4176                                         page->index,
4177                                         readahead_gfp_mask(mapping))) {
4178                         put_page(page);
4179                         continue;
4180                 }
4181
4182                 pagepool[nr++] = page;
4183                 if (nr < ARRAY_SIZE(pagepool))
4184                         continue;
4185                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4186                                    &bio, 0, &bio_flags, &prev_em_start);
4187                 nr = 0;
4188         }
4189         if (nr)
4190                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4191                                    &bio, 0, &bio_flags, &prev_em_start);
4192
4193         if (em_cached)
4194                 free_extent_map(em_cached);
4195
4196         BUG_ON(!list_empty(pages));
4197         if (bio)
4198                 return submit_one_bio(bio, 0, bio_flags);
4199         return 0;
4200 }
4201
4202 /*
4203  * basic invalidatepage code, this waits on any locked or writeback
4204  * ranges corresponding to the page, and then deletes any extent state
4205  * records from the tree
4206  */
4207 int extent_invalidatepage(struct extent_io_tree *tree,
4208                           struct page *page, unsigned long offset)
4209 {
4210         struct extent_state *cached_state = NULL;
4211         u64 start = page_offset(page);
4212         u64 end = start + PAGE_SIZE - 1;
4213         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4214
4215         start += ALIGN(offset, blocksize);
4216         if (start > end)
4217                 return 0;
4218
4219         lock_extent_bits(tree, start, end, &cached_state);
4220         wait_on_page_writeback(page);
4221         clear_extent_bit(tree, start, end,
4222                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4223                          EXTENT_DO_ACCOUNTING,
4224                          1, 1, &cached_state, GFP_NOFS);
4225         return 0;
4226 }
4227
4228 /*
4229  * a helper for releasepage, this tests for areas of the page that
4230  * are locked or under IO and drops the related state bits if it is safe
4231  * to drop the page.
4232  */
4233 static int try_release_extent_state(struct extent_map_tree *map,
4234                                     struct extent_io_tree *tree,
4235                                     struct page *page, gfp_t mask)
4236 {
4237         u64 start = page_offset(page);
4238         u64 end = start + PAGE_SIZE - 1;
4239         int ret = 1;
4240
4241         if (test_range_bit(tree, start, end,
4242                            EXTENT_IOBITS, 0, NULL))
4243                 ret = 0;
4244         else {
4245                 if ((mask & GFP_NOFS) == GFP_NOFS)
4246                         mask = GFP_NOFS;
4247                 /*
4248                  * at this point we can safely clear everything except the
4249                  * locked bit and the nodatasum bit
4250                  */
4251                 ret = clear_extent_bit(tree, start, end,
4252                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4253                                  0, 0, NULL, mask);
4254
4255                 /* if clear_extent_bit failed for enomem reasons,
4256                  * we can't allow the release to continue.
4257                  */
4258                 if (ret < 0)
4259                         ret = 0;
4260                 else
4261                         ret = 1;
4262         }
4263         return ret;
4264 }
4265
4266 /*
4267  * a helper for releasepage.  As long as there are no locked extents
4268  * in the range corresponding to the page, both state records and extent
4269  * map records are removed
4270  */
4271 int try_release_extent_mapping(struct extent_map_tree *map,
4272                                struct extent_io_tree *tree, struct page *page,
4273                                gfp_t mask)
4274 {
4275         struct extent_map *em;
4276         u64 start = page_offset(page);
4277         u64 end = start + PAGE_SIZE - 1;
4278
4279         if (gfpflags_allow_blocking(mask) &&
4280             page->mapping->host->i_size > SZ_16M) {
4281                 u64 len;
4282                 while (start <= end) {
4283                         len = end - start + 1;
4284                         write_lock(&map->lock);
4285                         em = lookup_extent_mapping(map, start, len);
4286                         if (!em) {
4287                                 write_unlock(&map->lock);
4288                                 break;
4289                         }
4290                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4291                             em->start != start) {
4292                                 write_unlock(&map->lock);
4293                                 free_extent_map(em);
4294                                 break;
4295                         }
4296                         if (!test_range_bit(tree, em->start,
4297                                             extent_map_end(em) - 1,
4298                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4299                                             0, NULL)) {
4300                                 remove_extent_mapping(map, em);
4301                                 /* once for the rb tree */
4302                                 free_extent_map(em);
4303                         }
4304                         start = extent_map_end(em);
4305                         write_unlock(&map->lock);
4306
4307                         /* once for us */
4308                         free_extent_map(em);
4309                 }
4310         }
4311         return try_release_extent_state(map, tree, page, mask);
4312 }
4313
4314 /*
4315  * helper function for fiemap, which doesn't want to see any holes.
4316  * This maps until we find something past 'last'
4317  */
4318 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4319                                                 u64 offset,
4320                                                 u64 last,
4321                                                 get_extent_t *get_extent)
4322 {
4323         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4324         struct extent_map *em;
4325         u64 len;
4326
4327         if (offset >= last)
4328                 return NULL;
4329
4330         while (1) {
4331                 len = last - offset;
4332                 if (len == 0)
4333                         break;
4334                 len = ALIGN(len, sectorsize);
4335                 em = get_extent(inode, NULL, 0, offset, len, 0);
4336                 if (IS_ERR_OR_NULL(em))
4337                         return em;
4338
4339                 /* if this isn't a hole return it */
4340                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4341                     em->block_start != EXTENT_MAP_HOLE) {
4342                         return em;
4343                 }
4344
4345                 /* this is a hole, advance to the next extent */
4346                 offset = extent_map_end(em);
4347                 free_extent_map(em);
4348                 if (offset >= last)
4349                         break;
4350         }
4351         return NULL;
4352 }
4353
4354 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4355                 __u64 start, __u64 len, get_extent_t *get_extent)
4356 {
4357         int ret = 0;
4358         u64 off = start;
4359         u64 max = start + len;
4360         u32 flags = 0;
4361         u32 found_type;
4362         u64 last;
4363         u64 last_for_get_extent = 0;
4364         u64 disko = 0;
4365         u64 isize = i_size_read(inode);
4366         struct btrfs_key found_key;
4367         struct extent_map *em = NULL;
4368         struct extent_state *cached_state = NULL;
4369         struct btrfs_path *path;
4370         struct btrfs_root *root = BTRFS_I(inode)->root;
4371         int end = 0;
4372         u64 em_start = 0;
4373         u64 em_len = 0;
4374         u64 em_end = 0;
4375
4376         if (len == 0)
4377                 return -EINVAL;
4378
4379         path = btrfs_alloc_path();
4380         if (!path)
4381                 return -ENOMEM;
4382         path->leave_spinning = 1;
4383
4384         start = round_down(start, BTRFS_I(inode)->root->sectorsize);
4385         len = round_up(max, BTRFS_I(inode)->root->sectorsize) - start;
4386
4387         /*
4388          * lookup the last file extent.  We're not using i_size here
4389          * because there might be preallocation past i_size
4390          */
4391         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode), -1,
4392                                        0);
4393         if (ret < 0) {
4394                 btrfs_free_path(path);
4395                 return ret;
4396         } else {
4397                 WARN_ON(!ret);
4398                 if (ret == 1)
4399                         ret = 0;
4400         }
4401
4402         path->slots[0]--;
4403         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4404         found_type = found_key.type;
4405
4406         /* No extents, but there might be delalloc bits */
4407         if (found_key.objectid != btrfs_ino(inode) ||
4408             found_type != BTRFS_EXTENT_DATA_KEY) {
4409                 /* have to trust i_size as the end */
4410                 last = (u64)-1;
4411                 last_for_get_extent = isize;
4412         } else {
4413                 /*
4414                  * remember the start of the last extent.  There are a
4415                  * bunch of different factors that go into the length of the
4416                  * extent, so its much less complex to remember where it started
4417                  */
4418                 last = found_key.offset;
4419                 last_for_get_extent = last + 1;
4420         }
4421         btrfs_release_path(path);
4422
4423         /*
4424          * we might have some extents allocated but more delalloc past those
4425          * extents.  so, we trust isize unless the start of the last extent is
4426          * beyond isize
4427          */
4428         if (last < isize) {
4429                 last = (u64)-1;
4430                 last_for_get_extent = isize;
4431         }
4432
4433         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4434                          &cached_state);
4435
4436         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4437                                    get_extent);
4438         if (!em)
4439                 goto out;
4440         if (IS_ERR(em)) {
4441                 ret = PTR_ERR(em);
4442                 goto out;
4443         }
4444
4445         while (!end) {
4446                 u64 offset_in_extent = 0;
4447
4448                 /* break if the extent we found is outside the range */
4449                 if (em->start >= max || extent_map_end(em) < off)
4450                         break;
4451
4452                 /*
4453                  * get_extent may return an extent that starts before our
4454                  * requested range.  We have to make sure the ranges
4455                  * we return to fiemap always move forward and don't
4456                  * overlap, so adjust the offsets here
4457                  */
4458                 em_start = max(em->start, off);
4459
4460                 /*
4461                  * record the offset from the start of the extent
4462                  * for adjusting the disk offset below.  Only do this if the
4463                  * extent isn't compressed since our in ram offset may be past
4464                  * what we have actually allocated on disk.
4465                  */
4466                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4467                         offset_in_extent = em_start - em->start;
4468                 em_end = extent_map_end(em);
4469                 em_len = em_end - em_start;
4470                 disko = 0;
4471                 flags = 0;
4472
4473                 /*
4474                  * bump off for our next call to get_extent
4475                  */
4476                 off = extent_map_end(em);
4477                 if (off >= max)
4478                         end = 1;
4479
4480                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4481                         end = 1;
4482                         flags |= FIEMAP_EXTENT_LAST;
4483                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4484                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4485                                   FIEMAP_EXTENT_NOT_ALIGNED);
4486                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4487                         flags |= (FIEMAP_EXTENT_DELALLOC |
4488                                   FIEMAP_EXTENT_UNKNOWN);
4489                 } else if (fieinfo->fi_extents_max) {
4490                         u64 bytenr = em->block_start -
4491                                 (em->start - em->orig_start);
4492
4493                         disko = em->block_start + offset_in_extent;
4494
4495                         /*
4496                          * As btrfs supports shared space, this information
4497                          * can be exported to userspace tools via
4498                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4499                          * then we're just getting a count and we can skip the
4500                          * lookup stuff.
4501                          */
4502                         ret = btrfs_check_shared(NULL, root->fs_info,
4503                                                  root->objectid,
4504                                                  btrfs_ino(inode), bytenr);
4505                         if (ret < 0)
4506                                 goto out_free;
4507                         if (ret)
4508                                 flags |= FIEMAP_EXTENT_SHARED;
4509                         ret = 0;
4510                 }
4511                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4512                         flags |= FIEMAP_EXTENT_ENCODED;
4513                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4514                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4515
4516                 free_extent_map(em);
4517                 em = NULL;
4518                 if ((em_start >= last) || em_len == (u64)-1 ||
4519                    (last == (u64)-1 && isize <= em_end)) {
4520                         flags |= FIEMAP_EXTENT_LAST;
4521                         end = 1;
4522                 }
4523
4524                 /* now scan forward to see if this is really the last extent. */
4525                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4526                                            get_extent);
4527                 if (IS_ERR(em)) {
4528                         ret = PTR_ERR(em);
4529                         goto out;
4530                 }
4531                 if (!em) {
4532                         flags |= FIEMAP_EXTENT_LAST;
4533                         end = 1;
4534                 }
4535                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4536                                               em_len, flags);
4537                 if (ret) {
4538                         if (ret == 1)
4539                                 ret = 0;
4540                         goto out_free;
4541                 }
4542         }
4543 out_free:
4544         free_extent_map(em);
4545 out:
4546         btrfs_free_path(path);
4547         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4548                              &cached_state, GFP_NOFS);
4549         return ret;
4550 }
4551
4552 static void __free_extent_buffer(struct extent_buffer *eb)
4553 {
4554         btrfs_leak_debug_del(&eb->leak_list);
4555         kmem_cache_free(extent_buffer_cache, eb);
4556 }
4557
4558 int extent_buffer_under_io(struct extent_buffer *eb)
4559 {
4560         return (atomic_read(&eb->io_pages) ||
4561                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4562                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4563 }
4564
4565 /*
4566  * Helper for releasing extent buffer page.
4567  */
4568 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4569 {
4570         unsigned long index;
4571         struct page *page;
4572         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4573
4574         BUG_ON(extent_buffer_under_io(eb));
4575
4576         index = num_extent_pages(eb->start, eb->len);
4577         if (index == 0)
4578                 return;
4579
4580         do {
4581                 index--;
4582                 page = eb->pages[index];
4583                 if (!page)
4584                         continue;
4585                 if (mapped)
4586                         spin_lock(&page->mapping->private_lock);
4587                 /*
4588                  * We do this since we'll remove the pages after we've
4589                  * removed the eb from the radix tree, so we could race
4590                  * and have this page now attached to the new eb.  So
4591                  * only clear page_private if it's still connected to
4592                  * this eb.
4593                  */
4594                 if (PagePrivate(page) &&
4595                     page->private == (unsigned long)eb) {
4596                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4597                         BUG_ON(PageDirty(page));
4598                         BUG_ON(PageWriteback(page));
4599                         /*
4600                          * We need to make sure we haven't be attached
4601                          * to a new eb.
4602                          */
4603                         ClearPagePrivate(page);
4604                         set_page_private(page, 0);
4605                         /* One for the page private */
4606                         put_page(page);
4607                 }
4608
4609                 if (mapped)
4610                         spin_unlock(&page->mapping->private_lock);
4611
4612                 /* One for when we allocated the page */
4613                 put_page(page);
4614         } while (index != 0);
4615 }
4616
4617 /*
4618  * Helper for releasing the extent buffer.
4619  */
4620 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4621 {
4622         btrfs_release_extent_buffer_page(eb);
4623         __free_extent_buffer(eb);
4624 }
4625
4626 static struct extent_buffer *
4627 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4628                       unsigned long len)
4629 {
4630         struct extent_buffer *eb = NULL;
4631
4632         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4633         eb->start = start;
4634         eb->len = len;
4635         eb->fs_info = fs_info;
4636         eb->bflags = 0;
4637         rwlock_init(&eb->lock);
4638         atomic_set(&eb->write_locks, 0);
4639         atomic_set(&eb->read_locks, 0);
4640         atomic_set(&eb->blocking_readers, 0);
4641         atomic_set(&eb->blocking_writers, 0);
4642         atomic_set(&eb->spinning_readers, 0);
4643         atomic_set(&eb->spinning_writers, 0);
4644         eb->lock_nested = 0;
4645         init_waitqueue_head(&eb->write_lock_wq);
4646         init_waitqueue_head(&eb->read_lock_wq);
4647
4648         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4649
4650         spin_lock_init(&eb->refs_lock);
4651         atomic_set(&eb->refs, 1);
4652         atomic_set(&eb->io_pages, 0);
4653
4654         /*
4655          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4656          */
4657         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4658                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4659         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4660
4661         return eb;
4662 }
4663
4664 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4665 {
4666         unsigned long i;
4667         struct page *p;
4668         struct extent_buffer *new;
4669         unsigned long num_pages = num_extent_pages(src->start, src->len);
4670
4671         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4672         if (new == NULL)
4673                 return NULL;
4674
4675         for (i = 0; i < num_pages; i++) {
4676                 p = alloc_page(GFP_NOFS);
4677                 if (!p) {
4678                         btrfs_release_extent_buffer(new);
4679                         return NULL;
4680                 }
4681                 attach_extent_buffer_page(new, p);
4682                 WARN_ON(PageDirty(p));
4683                 SetPageUptodate(p);
4684                 new->pages[i] = p;
4685         }
4686
4687         copy_extent_buffer(new, src, 0, 0, src->len);
4688         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4689         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4690
4691         return new;
4692 }
4693
4694 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4695                                                   u64 start, unsigned long len)
4696 {
4697         struct extent_buffer *eb;
4698         unsigned long num_pages;
4699         unsigned long i;
4700
4701         num_pages = num_extent_pages(start, len);
4702
4703         eb = __alloc_extent_buffer(fs_info, start, len);
4704         if (!eb)
4705                 return NULL;
4706
4707         for (i = 0; i < num_pages; i++) {
4708                 eb->pages[i] = alloc_page(GFP_NOFS);
4709                 if (!eb->pages[i])
4710                         goto err;
4711         }
4712         set_extent_buffer_uptodate(eb);
4713         btrfs_set_header_nritems(eb, 0);
4714         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4715
4716         return eb;
4717 err:
4718         for (; i > 0; i--)
4719                 __free_page(eb->pages[i - 1]);
4720         __free_extent_buffer(eb);
4721         return NULL;
4722 }
4723
4724 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4725                                                 u64 start, u32 nodesize)
4726 {
4727         unsigned long len;
4728
4729         if (!fs_info) {
4730                 /*
4731                  * Called only from tests that don't always have a fs_info
4732                  * available
4733                  */
4734                 len = nodesize;
4735         } else {
4736                 len = fs_info->tree_root->nodesize;
4737         }
4738
4739         return __alloc_dummy_extent_buffer(fs_info, start, len);
4740 }
4741
4742 static void check_buffer_tree_ref(struct extent_buffer *eb)
4743 {
4744         int refs;
4745         /* the ref bit is tricky.  We have to make sure it is set
4746          * if we have the buffer dirty.   Otherwise the
4747          * code to free a buffer can end up dropping a dirty
4748          * page
4749          *
4750          * Once the ref bit is set, it won't go away while the
4751          * buffer is dirty or in writeback, and it also won't
4752          * go away while we have the reference count on the
4753          * eb bumped.
4754          *
4755          * We can't just set the ref bit without bumping the
4756          * ref on the eb because free_extent_buffer might
4757          * see the ref bit and try to clear it.  If this happens
4758          * free_extent_buffer might end up dropping our original
4759          * ref by mistake and freeing the page before we are able
4760          * to add one more ref.
4761          *
4762          * So bump the ref count first, then set the bit.  If someone
4763          * beat us to it, drop the ref we added.
4764          */
4765         refs = atomic_read(&eb->refs);
4766         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4767                 return;
4768
4769         spin_lock(&eb->refs_lock);
4770         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4771                 atomic_inc(&eb->refs);
4772         spin_unlock(&eb->refs_lock);
4773 }
4774
4775 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4776                 struct page *accessed)
4777 {
4778         unsigned long num_pages, i;
4779
4780         check_buffer_tree_ref(eb);
4781
4782         num_pages = num_extent_pages(eb->start, eb->len);
4783         for (i = 0; i < num_pages; i++) {
4784                 struct page *p = eb->pages[i];
4785
4786                 if (p != accessed)
4787                         mark_page_accessed(p);
4788         }
4789 }
4790
4791 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4792                                          u64 start)
4793 {
4794         struct extent_buffer *eb;
4795
4796         rcu_read_lock();
4797         eb = radix_tree_lookup(&fs_info->buffer_radix,
4798                                start >> PAGE_SHIFT);
4799         if (eb && atomic_inc_not_zero(&eb->refs)) {
4800                 rcu_read_unlock();
4801                 /*
4802                  * Lock our eb's refs_lock to avoid races with
4803                  * free_extent_buffer. When we get our eb it might be flagged
4804                  * with EXTENT_BUFFER_STALE and another task running
4805                  * free_extent_buffer might have seen that flag set,
4806                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4807                  * writeback flags not set) and it's still in the tree (flag
4808                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4809                  * of decrementing the extent buffer's reference count twice.
4810                  * So here we could race and increment the eb's reference count,
4811                  * clear its stale flag, mark it as dirty and drop our reference
4812                  * before the other task finishes executing free_extent_buffer,
4813                  * which would later result in an attempt to free an extent
4814                  * buffer that is dirty.
4815                  */
4816                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4817                         spin_lock(&eb->refs_lock);
4818                         spin_unlock(&eb->refs_lock);
4819                 }
4820                 mark_extent_buffer_accessed(eb, NULL);
4821                 return eb;
4822         }
4823         rcu_read_unlock();
4824
4825         return NULL;
4826 }
4827
4828 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4829 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4830                                         u64 start, u32 nodesize)
4831 {
4832         struct extent_buffer *eb, *exists = NULL;
4833         int ret;
4834
4835         eb = find_extent_buffer(fs_info, start);
4836         if (eb)
4837                 return eb;
4838         eb = alloc_dummy_extent_buffer(fs_info, start, nodesize);
4839         if (!eb)
4840                 return NULL;
4841         eb->fs_info = fs_info;
4842 again:
4843         ret = radix_tree_preload(GFP_NOFS);
4844         if (ret)
4845                 goto free_eb;
4846         spin_lock(&fs_info->buffer_lock);
4847         ret = radix_tree_insert(&fs_info->buffer_radix,
4848                                 start >> PAGE_SHIFT, eb);
4849         spin_unlock(&fs_info->buffer_lock);
4850         radix_tree_preload_end();
4851         if (ret == -EEXIST) {
4852                 exists = find_extent_buffer(fs_info, start);
4853                 if (exists)
4854                         goto free_eb;
4855                 else
4856                         goto again;
4857         }
4858         check_buffer_tree_ref(eb);
4859         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4860
4861         /*
4862          * We will free dummy extent buffer's if they come into
4863          * free_extent_buffer with a ref count of 2, but if we are using this we
4864          * want the buffers to stay in memory until we're done with them, so
4865          * bump the ref count again.
4866          */
4867         atomic_inc(&eb->refs);
4868         return eb;
4869 free_eb:
4870         btrfs_release_extent_buffer(eb);
4871         return exists;
4872 }
4873 #endif
4874
4875 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4876                                           u64 start)
4877 {
4878         unsigned long len = fs_info->tree_root->nodesize;
4879         unsigned long num_pages = num_extent_pages(start, len);
4880         unsigned long i;
4881         unsigned long index = start >> PAGE_SHIFT;
4882         struct extent_buffer *eb;
4883         struct extent_buffer *exists = NULL;
4884         struct page *p;
4885         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4886         int uptodate = 1;
4887         int ret;
4888
4889         if (!IS_ALIGNED(start, fs_info->tree_root->sectorsize)) {
4890                 btrfs_err(fs_info, "bad tree block start %llu", start);
4891                 return ERR_PTR(-EINVAL);
4892         }
4893
4894         eb = find_extent_buffer(fs_info, start);
4895         if (eb)
4896                 return eb;
4897
4898         eb = __alloc_extent_buffer(fs_info, start, len);
4899         if (!eb)
4900                 return ERR_PTR(-ENOMEM);
4901
4902         for (i = 0; i < num_pages; i++, index++) {
4903                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4904                 if (!p) {
4905                         exists = ERR_PTR(-ENOMEM);
4906                         goto free_eb;
4907                 }
4908
4909                 spin_lock(&mapping->private_lock);
4910                 if (PagePrivate(p)) {
4911                         /*
4912                          * We could have already allocated an eb for this page
4913                          * and attached one so lets see if we can get a ref on
4914                          * the existing eb, and if we can we know it's good and
4915                          * we can just return that one, else we know we can just
4916                          * overwrite page->private.
4917                          */
4918                         exists = (struct extent_buffer *)p->private;
4919                         if (atomic_inc_not_zero(&exists->refs)) {
4920                                 spin_unlock(&mapping->private_lock);
4921                                 unlock_page(p);
4922                                 put_page(p);
4923                                 mark_extent_buffer_accessed(exists, p);
4924                                 goto free_eb;
4925                         }
4926                         exists = NULL;
4927
4928                         /*
4929                          * Do this so attach doesn't complain and we need to
4930                          * drop the ref the old guy had.
4931                          */
4932                         ClearPagePrivate(p);
4933                         WARN_ON(PageDirty(p));
4934                         put_page(p);
4935                 }
4936                 attach_extent_buffer_page(eb, p);
4937                 spin_unlock(&mapping->private_lock);
4938                 WARN_ON(PageDirty(p));
4939                 eb->pages[i] = p;
4940                 if (!PageUptodate(p))
4941                         uptodate = 0;
4942
4943                 /*
4944                  * see below about how we avoid a nasty race with release page
4945                  * and why we unlock later
4946                  */
4947         }
4948         if (uptodate)
4949                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4950 again:
4951         ret = radix_tree_preload(GFP_NOFS);
4952         if (ret) {
4953                 exists = ERR_PTR(ret);
4954                 goto free_eb;
4955         }
4956
4957         spin_lock(&fs_info->buffer_lock);
4958         ret = radix_tree_insert(&fs_info->buffer_radix,
4959                                 start >> PAGE_SHIFT, eb);
4960         spin_unlock(&fs_info->buffer_lock);
4961         radix_tree_preload_end();
4962         if (ret == -EEXIST) {
4963                 exists = find_extent_buffer(fs_info, start);
4964                 if (exists)
4965                         goto free_eb;
4966                 else
4967                         goto again;
4968         }
4969         /* add one reference for the tree */
4970         check_buffer_tree_ref(eb);
4971         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4972
4973         /*
4974          * there is a race where release page may have
4975          * tried to find this extent buffer in the radix
4976          * but failed.  It will tell the VM it is safe to
4977          * reclaim the, and it will clear the page private bit.
4978          * We must make sure to set the page private bit properly
4979          * after the extent buffer is in the radix tree so
4980          * it doesn't get lost
4981          */
4982         SetPageChecked(eb->pages[0]);
4983         for (i = 1; i < num_pages; i++) {
4984                 p = eb->pages[i];
4985                 ClearPageChecked(p);
4986                 unlock_page(p);
4987         }
4988         unlock_page(eb->pages[0]);
4989         return eb;
4990
4991 free_eb:
4992         WARN_ON(!atomic_dec_and_test(&eb->refs));
4993         for (i = 0; i < num_pages; i++) {
4994                 if (eb->pages[i])
4995                         unlock_page(eb->pages[i]);
4996         }
4997
4998         btrfs_release_extent_buffer(eb);
4999         return exists;
5000 }
5001
5002 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5003 {
5004         struct extent_buffer *eb =
5005                         container_of(head, struct extent_buffer, rcu_head);
5006
5007         __free_extent_buffer(eb);
5008 }
5009
5010 /* Expects to have eb->eb_lock already held */
5011 static int release_extent_buffer(struct extent_buffer *eb)
5012 {
5013         WARN_ON(atomic_read(&eb->refs) == 0);
5014         if (atomic_dec_and_test(&eb->refs)) {
5015                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5016                         struct btrfs_fs_info *fs_info = eb->fs_info;
5017
5018                         spin_unlock(&eb->refs_lock);
5019
5020                         spin_lock(&fs_info->buffer_lock);
5021                         radix_tree_delete(&fs_info->buffer_radix,
5022                                           eb->start >> PAGE_SHIFT);
5023                         spin_unlock(&fs_info->buffer_lock);
5024                 } else {
5025                         spin_unlock(&eb->refs_lock);
5026                 }
5027
5028                 /* Should be safe to release our pages at this point */
5029                 btrfs_release_extent_buffer_page(eb);
5030 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5031                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5032                         __free_extent_buffer(eb);
5033                         return 1;
5034                 }
5035 #endif
5036                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5037                 return 1;
5038         }
5039         spin_unlock(&eb->refs_lock);
5040
5041         return 0;
5042 }
5043
5044 void free_extent_buffer(struct extent_buffer *eb)
5045 {
5046         int refs;
5047         int old;
5048         if (!eb)
5049                 return;
5050
5051         while (1) {
5052                 refs = atomic_read(&eb->refs);
5053                 if (refs <= 3)
5054                         break;
5055                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5056                 if (old == refs)
5057                         return;
5058         }
5059
5060         spin_lock(&eb->refs_lock);
5061         if (atomic_read(&eb->refs) == 2 &&
5062             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5063                 atomic_dec(&eb->refs);
5064
5065         if (atomic_read(&eb->refs) == 2 &&
5066             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5067             !extent_buffer_under_io(eb) &&
5068             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5069                 atomic_dec(&eb->refs);
5070
5071         /*
5072          * I know this is terrible, but it's temporary until we stop tracking
5073          * the uptodate bits and such for the extent buffers.
5074          */
5075         release_extent_buffer(eb);
5076 }
5077
5078 void free_extent_buffer_stale(struct extent_buffer *eb)
5079 {
5080         if (!eb)
5081                 return;
5082
5083         spin_lock(&eb->refs_lock);
5084         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5085
5086         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5087             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5088                 atomic_dec(&eb->refs);
5089         release_extent_buffer(eb);
5090 }
5091
5092 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5093 {
5094         unsigned long i;
5095         unsigned long num_pages;
5096         struct page *page;
5097
5098         num_pages = num_extent_pages(eb->start, eb->len);
5099
5100         for (i = 0; i < num_pages; i++) {
5101                 page = eb->pages[i];
5102                 if (!PageDirty(page))
5103                         continue;
5104
5105                 lock_page(page);
5106                 WARN_ON(!PagePrivate(page));
5107
5108                 clear_page_dirty_for_io(page);
5109                 spin_lock_irq(&page->mapping->tree_lock);
5110                 if (!PageDirty(page)) {
5111                         radix_tree_tag_clear(&page->mapping->page_tree,
5112                                                 page_index(page),
5113                                                 PAGECACHE_TAG_DIRTY);
5114                 }
5115                 spin_unlock_irq(&page->mapping->tree_lock);
5116                 ClearPageError(page);
5117                 unlock_page(page);
5118         }
5119         WARN_ON(atomic_read(&eb->refs) == 0);
5120 }
5121
5122 int set_extent_buffer_dirty(struct extent_buffer *eb)
5123 {
5124         unsigned long i;
5125         unsigned long num_pages;
5126         int was_dirty = 0;
5127
5128         check_buffer_tree_ref(eb);
5129
5130         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5131
5132         num_pages = num_extent_pages(eb->start, eb->len);
5133         WARN_ON(atomic_read(&eb->refs) == 0);
5134         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5135
5136         for (i = 0; i < num_pages; i++)
5137                 set_page_dirty(eb->pages[i]);
5138         return was_dirty;
5139 }
5140
5141 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5142 {
5143         unsigned long i;
5144         struct page *page;
5145         unsigned long num_pages;
5146
5147         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5148         num_pages = num_extent_pages(eb->start, eb->len);
5149         for (i = 0; i < num_pages; i++) {
5150                 page = eb->pages[i];
5151                 if (page)
5152                         ClearPageUptodate(page);
5153         }
5154 }
5155
5156 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5157 {
5158         unsigned long i;
5159         struct page *page;
5160         unsigned long num_pages;
5161
5162         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5163         num_pages = num_extent_pages(eb->start, eb->len);
5164         for (i = 0; i < num_pages; i++) {
5165                 page = eb->pages[i];
5166                 SetPageUptodate(page);
5167         }
5168 }
5169
5170 int extent_buffer_uptodate(struct extent_buffer *eb)
5171 {
5172         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5173 }
5174
5175 int read_extent_buffer_pages(struct extent_io_tree *tree,
5176                              struct extent_buffer *eb, u64 start, int wait,
5177                              get_extent_t *get_extent, int mirror_num)
5178 {
5179         unsigned long i;
5180         unsigned long start_i;
5181         struct page *page;
5182         int err;
5183         int ret = 0;
5184         int locked_pages = 0;
5185         int all_uptodate = 1;
5186         unsigned long num_pages;
5187         unsigned long num_reads = 0;
5188         struct bio *bio = NULL;
5189         unsigned long bio_flags = 0;
5190
5191         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5192                 return 0;
5193
5194         if (start) {
5195                 WARN_ON(start < eb->start);
5196                 start_i = (start >> PAGE_SHIFT) -
5197                         (eb->start >> PAGE_SHIFT);
5198         } else {
5199                 start_i = 0;
5200         }
5201
5202         num_pages = num_extent_pages(eb->start, eb->len);
5203         for (i = start_i; i < num_pages; i++) {
5204                 page = eb->pages[i];
5205                 if (wait == WAIT_NONE) {
5206                         if (!trylock_page(page))
5207                                 goto unlock_exit;
5208                 } else {
5209                         lock_page(page);
5210                 }
5211                 locked_pages++;
5212                 if (!PageUptodate(page)) {
5213                         num_reads++;
5214                         all_uptodate = 0;
5215                 }
5216         }
5217         if (all_uptodate) {
5218                 if (start_i == 0)
5219                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5220                 goto unlock_exit;
5221         }
5222
5223         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5224         eb->read_mirror = 0;
5225         atomic_set(&eb->io_pages, num_reads);
5226         for (i = start_i; i < num_pages; i++) {
5227                 page = eb->pages[i];
5228
5229                 if (!PageUptodate(page)) {
5230                         if (ret) {
5231                                 atomic_dec(&eb->io_pages);
5232                                 unlock_page(page);
5233                                 continue;
5234                         }
5235
5236                         ClearPageError(page);
5237                         err = __extent_read_full_page(tree, page,
5238                                                       get_extent, &bio,
5239                                                       mirror_num, &bio_flags,
5240                                                       REQ_META);
5241                         if (err) {
5242                                 ret = err;
5243                                 /*
5244                                  * We use &bio in above __extent_read_full_page,
5245                                  * so we ensure that if it returns error, the
5246                                  * current page fails to add itself to bio and
5247                                  * it's been unlocked.
5248                                  *
5249                                  * We must dec io_pages by ourselves.
5250                                  */
5251                                 atomic_dec(&eb->io_pages);
5252                         }
5253                 } else {
5254                         unlock_page(page);
5255                 }
5256         }
5257
5258         if (bio) {
5259                 err = submit_one_bio(bio, mirror_num, bio_flags);
5260                 if (err)
5261                         return err;
5262         }
5263
5264         if (ret || wait != WAIT_COMPLETE)
5265                 return ret;
5266
5267         for (i = start_i; i < num_pages; i++) {
5268                 page = eb->pages[i];
5269                 wait_on_page_locked(page);
5270                 if (!PageUptodate(page))
5271                         ret = -EIO;
5272         }
5273
5274         return ret;
5275
5276 unlock_exit:
5277         i = start_i;
5278         while (locked_pages > 0) {
5279                 page = eb->pages[i];
5280                 i++;
5281                 unlock_page(page);
5282                 locked_pages--;
5283         }
5284         return ret;
5285 }
5286
5287 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5288                         unsigned long start,
5289                         unsigned long len)
5290 {
5291         size_t cur;
5292         size_t offset;
5293         struct page *page;
5294         char *kaddr;
5295         char *dst = (char *)dstv;
5296         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5297         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5298
5299         WARN_ON(start > eb->len);
5300         WARN_ON(start + len > eb->start + eb->len);
5301
5302         offset = (start_offset + start) & (PAGE_SIZE - 1);
5303
5304         while (len > 0) {
5305                 page = eb->pages[i];
5306
5307                 cur = min(len, (PAGE_SIZE - offset));
5308                 kaddr = page_address(page);
5309                 memcpy(dst, kaddr + offset, cur);
5310
5311                 dst += cur;
5312                 len -= cur;
5313                 offset = 0;
5314                 i++;
5315         }
5316 }
5317
5318 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5319                         unsigned long start,
5320                         unsigned long len)
5321 {
5322         size_t cur;
5323         size_t offset;
5324         struct page *page;
5325         char *kaddr;
5326         char __user *dst = (char __user *)dstv;
5327         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5328         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5329         int ret = 0;
5330
5331         WARN_ON(start > eb->len);
5332         WARN_ON(start + len > eb->start + eb->len);
5333
5334         offset = (start_offset + start) & (PAGE_SIZE - 1);
5335
5336         while (len > 0) {
5337                 page = eb->pages[i];
5338
5339                 cur = min(len, (PAGE_SIZE - offset));
5340                 kaddr = page_address(page);
5341                 if (copy_to_user(dst, kaddr + offset, cur)) {
5342                         ret = -EFAULT;
5343                         break;
5344                 }
5345
5346                 dst += cur;
5347                 len -= cur;
5348                 offset = 0;
5349                 i++;
5350         }
5351
5352         return ret;
5353 }
5354
5355 /*
5356  * return 0 if the item is found within a page.
5357  * return 1 if the item spans two pages.
5358  * return -EINVAL otherwise.
5359  */
5360 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5361                                unsigned long min_len, char **map,
5362                                unsigned long *map_start,
5363                                unsigned long *map_len)
5364 {
5365         size_t offset = start & (PAGE_SIZE - 1);
5366         char *kaddr;
5367         struct page *p;
5368         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5369         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5370         unsigned long end_i = (start_offset + start + min_len - 1) >>
5371                 PAGE_SHIFT;
5372
5373         if (i != end_i)
5374                 return 1;
5375
5376         if (i == 0) {
5377                 offset = start_offset;
5378                 *map_start = 0;
5379         } else {
5380                 offset = 0;
5381                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5382         }
5383
5384         if (start + min_len > eb->len) {
5385                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
5386                        "wanted %lu %lu\n",
5387                        eb->start, eb->len, start, min_len);
5388                 return -EINVAL;
5389         }
5390
5391         p = eb->pages[i];
5392         kaddr = page_address(p);
5393         *map = kaddr + offset;
5394         *map_len = PAGE_SIZE - offset;
5395         return 0;
5396 }
5397
5398 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5399                           unsigned long start,
5400                           unsigned long len)
5401 {
5402         size_t cur;
5403         size_t offset;
5404         struct page *page;
5405         char *kaddr;
5406         char *ptr = (char *)ptrv;
5407         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5408         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5409         int ret = 0;
5410
5411         WARN_ON(start > eb->len);
5412         WARN_ON(start + len > eb->start + eb->len);
5413
5414         offset = (start_offset + start) & (PAGE_SIZE - 1);
5415
5416         while (len > 0) {
5417                 page = eb->pages[i];
5418
5419                 cur = min(len, (PAGE_SIZE - offset));
5420
5421                 kaddr = page_address(page);
5422                 ret = memcmp(ptr, kaddr + offset, cur);
5423                 if (ret)
5424                         break;
5425
5426                 ptr += cur;
5427                 len -= cur;
5428                 offset = 0;
5429                 i++;
5430         }
5431         return ret;
5432 }
5433
5434 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5435                          unsigned long start, unsigned long len)
5436 {
5437         size_t cur;
5438         size_t offset;
5439         struct page *page;
5440         char *kaddr;
5441         char *src = (char *)srcv;
5442         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5443         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5444
5445         WARN_ON(start > eb->len);
5446         WARN_ON(start + len > eb->start + eb->len);
5447
5448         offset = (start_offset + start) & (PAGE_SIZE - 1);
5449
5450         while (len > 0) {
5451                 page = eb->pages[i];
5452                 WARN_ON(!PageUptodate(page));
5453
5454                 cur = min(len, PAGE_SIZE - offset);
5455                 kaddr = page_address(page);
5456                 memcpy(kaddr + offset, src, cur);
5457
5458                 src += cur;
5459                 len -= cur;
5460                 offset = 0;
5461                 i++;
5462         }
5463 }
5464
5465 void memset_extent_buffer(struct extent_buffer *eb, char c,
5466                           unsigned long start, unsigned long len)
5467 {
5468         size_t cur;
5469         size_t offset;
5470         struct page *page;
5471         char *kaddr;
5472         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5473         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5474
5475         WARN_ON(start > eb->len);
5476         WARN_ON(start + len > eb->start + eb->len);
5477
5478         offset = (start_offset + start) & (PAGE_SIZE - 1);
5479
5480         while (len > 0) {
5481                 page = eb->pages[i];
5482                 WARN_ON(!PageUptodate(page));
5483
5484                 cur = min(len, PAGE_SIZE - offset);
5485                 kaddr = page_address(page);
5486                 memset(kaddr + offset, c, cur);
5487
5488                 len -= cur;
5489                 offset = 0;
5490                 i++;
5491         }
5492 }
5493
5494 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5495                         unsigned long dst_offset, unsigned long src_offset,
5496                         unsigned long len)
5497 {
5498         u64 dst_len = dst->len;
5499         size_t cur;
5500         size_t offset;
5501         struct page *page;
5502         char *kaddr;
5503         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5504         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5505
5506         WARN_ON(src->len != dst_len);
5507
5508         offset = (start_offset + dst_offset) &
5509                 (PAGE_SIZE - 1);
5510
5511         while (len > 0) {
5512                 page = dst->pages[i];
5513                 WARN_ON(!PageUptodate(page));
5514
5515                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5516
5517                 kaddr = page_address(page);
5518                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5519
5520                 src_offset += cur;
5521                 len -= cur;
5522                 offset = 0;
5523                 i++;
5524         }
5525 }
5526
5527 /*
5528  * The extent buffer bitmap operations are done with byte granularity because
5529  * bitmap items are not guaranteed to be aligned to a word and therefore a
5530  * single word in a bitmap may straddle two pages in the extent buffer.
5531  */
5532 #define BIT_BYTE(nr) ((nr) / BITS_PER_BYTE)
5533 #define BYTE_MASK ((1 << BITS_PER_BYTE) - 1)
5534 #define BITMAP_FIRST_BYTE_MASK(start) \
5535         ((BYTE_MASK << ((start) & (BITS_PER_BYTE - 1))) & BYTE_MASK)
5536 #define BITMAP_LAST_BYTE_MASK(nbits) \
5537         (BYTE_MASK >> (-(nbits) & (BITS_PER_BYTE - 1)))
5538
5539 /*
5540  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5541  * given bit number
5542  * @eb: the extent buffer
5543  * @start: offset of the bitmap item in the extent buffer
5544  * @nr: bit number
5545  * @page_index: return index of the page in the extent buffer that contains the
5546  * given bit number
5547  * @page_offset: return offset into the page given by page_index
5548  *
5549  * This helper hides the ugliness of finding the byte in an extent buffer which
5550  * contains a given bit.
5551  */
5552 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5553                                     unsigned long start, unsigned long nr,
5554                                     unsigned long *page_index,
5555                                     size_t *page_offset)
5556 {
5557         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5558         size_t byte_offset = BIT_BYTE(nr);
5559         size_t offset;
5560
5561         /*
5562          * The byte we want is the offset of the extent buffer + the offset of
5563          * the bitmap item in the extent buffer + the offset of the byte in the
5564          * bitmap item.
5565          */
5566         offset = start_offset + start + byte_offset;
5567
5568         *page_index = offset >> PAGE_SHIFT;
5569         *page_offset = offset & (PAGE_SIZE - 1);
5570 }
5571
5572 /**
5573  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5574  * @eb: the extent buffer
5575  * @start: offset of the bitmap item in the extent buffer
5576  * @nr: bit number to test
5577  */
5578 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5579                            unsigned long nr)
5580 {
5581         char *kaddr;
5582         struct page *page;
5583         unsigned long i;
5584         size_t offset;
5585
5586         eb_bitmap_offset(eb, start, nr, &i, &offset);
5587         page = eb->pages[i];
5588         WARN_ON(!PageUptodate(page));
5589         kaddr = page_address(page);
5590         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5591 }
5592
5593 /**
5594  * extent_buffer_bitmap_set - set an area of a bitmap
5595  * @eb: the extent buffer
5596  * @start: offset of the bitmap item in the extent buffer
5597  * @pos: bit number of the first bit
5598  * @len: number of bits to set
5599  */
5600 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5601                               unsigned long pos, unsigned long len)
5602 {
5603         char *kaddr;
5604         struct page *page;
5605         unsigned long i;
5606         size_t offset;
5607         const unsigned int size = pos + len;
5608         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5609         unsigned int mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5610
5611         eb_bitmap_offset(eb, start, pos, &i, &offset);
5612         page = eb->pages[i];
5613         WARN_ON(!PageUptodate(page));
5614         kaddr = page_address(page);
5615
5616         while (len >= bits_to_set) {
5617                 kaddr[offset] |= mask_to_set;
5618                 len -= bits_to_set;
5619                 bits_to_set = BITS_PER_BYTE;
5620                 mask_to_set = ~0U;
5621                 if (++offset >= PAGE_SIZE && len > 0) {
5622                         offset = 0;
5623                         page = eb->pages[++i];
5624                         WARN_ON(!PageUptodate(page));
5625                         kaddr = page_address(page);
5626                 }
5627         }
5628         if (len) {
5629                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5630                 kaddr[offset] |= mask_to_set;
5631         }
5632 }
5633
5634
5635 /**
5636  * extent_buffer_bitmap_clear - clear an area of a bitmap
5637  * @eb: the extent buffer
5638  * @start: offset of the bitmap item in the extent buffer
5639  * @pos: bit number of the first bit
5640  * @len: number of bits to clear
5641  */
5642 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5643                                 unsigned long pos, unsigned long len)
5644 {
5645         char *kaddr;
5646         struct page *page;
5647         unsigned long i;
5648         size_t offset;
5649         const unsigned int size = pos + len;
5650         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5651         unsigned int mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5652
5653         eb_bitmap_offset(eb, start, pos, &i, &offset);
5654         page = eb->pages[i];
5655         WARN_ON(!PageUptodate(page));
5656         kaddr = page_address(page);
5657
5658         while (len >= bits_to_clear) {
5659                 kaddr[offset] &= ~mask_to_clear;
5660                 len -= bits_to_clear;
5661                 bits_to_clear = BITS_PER_BYTE;
5662                 mask_to_clear = ~0U;
5663                 if (++offset >= PAGE_SIZE && len > 0) {
5664                         offset = 0;
5665                         page = eb->pages[++i];
5666                         WARN_ON(!PageUptodate(page));
5667                         kaddr = page_address(page);
5668                 }
5669         }
5670         if (len) {
5671                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5672                 kaddr[offset] &= ~mask_to_clear;
5673         }
5674 }
5675
5676 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5677 {
5678         unsigned long distance = (src > dst) ? src - dst : dst - src;
5679         return distance < len;
5680 }
5681
5682 static void copy_pages(struct page *dst_page, struct page *src_page,
5683                        unsigned long dst_off, unsigned long src_off,
5684                        unsigned long len)
5685 {
5686         char *dst_kaddr = page_address(dst_page);
5687         char *src_kaddr;
5688         int must_memmove = 0;
5689
5690         if (dst_page != src_page) {
5691                 src_kaddr = page_address(src_page);
5692         } else {
5693                 src_kaddr = dst_kaddr;
5694                 if (areas_overlap(src_off, dst_off, len))
5695                         must_memmove = 1;
5696         }
5697
5698         if (must_memmove)
5699                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5700         else
5701                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5702 }
5703
5704 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5705                            unsigned long src_offset, unsigned long len)
5706 {
5707         size_t cur;
5708         size_t dst_off_in_page;
5709         size_t src_off_in_page;
5710         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5711         unsigned long dst_i;
5712         unsigned long src_i;
5713
5714         if (src_offset + len > dst->len) {
5715                 btrfs_err(dst->fs_info,
5716                         "memmove bogus src_offset %lu move "
5717                        "len %lu dst len %lu", src_offset, len, dst->len);
5718                 BUG_ON(1);
5719         }
5720         if (dst_offset + len > dst->len) {
5721                 btrfs_err(dst->fs_info,
5722                         "memmove bogus dst_offset %lu move "
5723                        "len %lu dst len %lu", dst_offset, len, dst->len);
5724                 BUG_ON(1);
5725         }
5726
5727         while (len > 0) {
5728                 dst_off_in_page = (start_offset + dst_offset) &
5729                         (PAGE_SIZE - 1);
5730                 src_off_in_page = (start_offset + src_offset) &
5731                         (PAGE_SIZE - 1);
5732
5733                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5734                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5735
5736                 cur = min(len, (unsigned long)(PAGE_SIZE -
5737                                                src_off_in_page));
5738                 cur = min_t(unsigned long, cur,
5739                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5740
5741                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5742                            dst_off_in_page, src_off_in_page, cur);
5743
5744                 src_offset += cur;
5745                 dst_offset += cur;
5746                 len -= cur;
5747         }
5748 }
5749
5750 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5751                            unsigned long src_offset, unsigned long len)
5752 {
5753         size_t cur;
5754         size_t dst_off_in_page;
5755         size_t src_off_in_page;
5756         unsigned long dst_end = dst_offset + len - 1;
5757         unsigned long src_end = src_offset + len - 1;
5758         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5759         unsigned long dst_i;
5760         unsigned long src_i;
5761
5762         if (src_offset + len > dst->len) {
5763                 btrfs_err(dst->fs_info, "memmove bogus src_offset %lu move "
5764                        "len %lu len %lu", src_offset, len, dst->len);
5765                 BUG_ON(1);
5766         }
5767         if (dst_offset + len > dst->len) {
5768                 btrfs_err(dst->fs_info, "memmove bogus dst_offset %lu move "
5769                        "len %lu len %lu", dst_offset, len, dst->len);
5770                 BUG_ON(1);
5771         }
5772         if (dst_offset < src_offset) {
5773                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5774                 return;
5775         }
5776         while (len > 0) {
5777                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5778                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5779
5780                 dst_off_in_page = (start_offset + dst_end) &
5781                         (PAGE_SIZE - 1);
5782                 src_off_in_page = (start_offset + src_end) &
5783                         (PAGE_SIZE - 1);
5784
5785                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5786                 cur = min(cur, dst_off_in_page + 1);
5787                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5788                            dst_off_in_page - cur + 1,
5789                            src_off_in_page - cur + 1, cur);
5790
5791                 dst_end -= cur;
5792                 src_end -= cur;
5793                 len -= cur;
5794         }
5795 }
5796
5797 int try_release_extent_buffer(struct page *page)
5798 {
5799         struct extent_buffer *eb;
5800
5801         /*
5802          * We need to make sure nobody is attaching this page to an eb right
5803          * now.
5804          */
5805         spin_lock(&page->mapping->private_lock);
5806         if (!PagePrivate(page)) {
5807                 spin_unlock(&page->mapping->private_lock);
5808                 return 1;
5809         }
5810
5811         eb = (struct extent_buffer *)page->private;
5812         BUG_ON(!eb);
5813
5814         /*
5815          * This is a little awful but should be ok, we need to make sure that
5816          * the eb doesn't disappear out from under us while we're looking at
5817          * this page.
5818          */
5819         spin_lock(&eb->refs_lock);
5820         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5821                 spin_unlock(&eb->refs_lock);
5822                 spin_unlock(&page->mapping->private_lock);
5823                 return 0;
5824         }
5825         spin_unlock(&page->mapping->private_lock);
5826
5827         /*
5828          * If tree ref isn't set then we know the ref on this eb is a real ref,
5829          * so just return, this page will likely be freed soon anyway.
5830          */
5831         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5832                 spin_unlock(&eb->refs_lock);
5833                 return 0;
5834         }
5835
5836         return release_extent_buffer(eb);
5837 }