fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/spinlock.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/swap.h>
  10 #include <linux/writeback.h>
  11 #include <linux/pagevec.h>
  12 #include <linux/prefetch.h>
  13 #include <linux/cleancache.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "ctree.h"
  17 #include "btrfs_inode.h"
  18 #include "volumes.h"
  19 #include "check-integrity.h"
  20 #include "locking.h"
  21 #include "rcu-string.h"
  22 #include "backref.h"
  23
  24 static struct kmem_cache *extent_state_cache;
  25 static struct kmem_cache *extent_buffer_cache;
  26 static struct bio_set *btrfs_bioset;
  27
  28 static inline bool extent_state_in_tree(const struct extent_state *state)
  29 {
  30         return !RB_EMPTY_NODE(&state->rb_node);
  31 }
  32
  33 #ifdef CONFIG_BTRFS_DEBUG
  34 static LIST_HEAD(buffers);
  35 static LIST_HEAD(states);
  36
  37 static DEFINE_SPINLOCK(leak_lock);
  38
  39 static inline
  40 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  41 {
  42         unsigned long flags;
  43
  44         spin_lock_irqsave(&leak_lock, flags);
  45         list_add(new, head);
  46         spin_unlock_irqrestore(&leak_lock, flags);
  47 }
  48
  49 static inline
  50 void btrfs_leak_debug_del(struct list_head *entry)
  51 {
  52         unsigned long flags;
  53
  54         spin_lock_irqsave(&leak_lock, flags);
  55         list_del(entry);
  56         spin_unlock_irqrestore(&leak_lock, flags);
  57 }
  58
  59 static inline
  60 void btrfs_leak_debug_check(void)
  61 {
  62         struct extent_state *state;
  63         struct extent_buffer *eb;
  64
  65         while (!list_empty(&states)) {
  66                 state = list_entry(states.next, struct extent_state, leak_list);
  67                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  68                        state->start, state->end, state->state,
  69                        extent_state_in_tree(state),
  70                        atomic_read(&state->refs));
  71                 list_del(&state->leak_list);
  72                 kmem_cache_free(extent_state_cache, state);
  73         }
  74
  75         while (!list_empty(&buffers)) {
  76                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  77                 printk(KERN_ERR "BTRFS: buffer leak start %llu len %lu "
  78                        "refs %d\n",
  79                        eb->start, eb->len, atomic_read(&eb->refs));
  80                 list_del(&eb->leak_list);
  81                 kmem_cache_free(extent_buffer_cache, eb);
  82         }
  83 }
  84
  85 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  86         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  87 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  88                 struct extent_io_tree *tree, u64 start, u64 end)
  89 {
  90         struct inode *inode;
  91         u64 isize;
  92
  93         if (!tree->mapping)
  94                 return;
  95
  96         inode = tree->mapping->host;
  97         isize = i_size_read(inode);
  98         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
  99                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
 100                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
 101                                 caller, btrfs_ino(inode), isize, start, end);
 102         }
 103 }
 104 #else
 105 #define btrfs_leak_debug_add(new, head) do {} while (0)
 106 #define btrfs_leak_debug_del(entry)     do {} while (0)
 107 #define btrfs_leak_debug_check()        do {} while (0)
 108 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 109 #endif
 110
 111 #define BUFFER_LRU_MAX 64
 112
 113 struct tree_entry {
 114         u64 start;
 115         u64 end;
 116         struct rb_node rb_node;
 117 };
 118
 119 struct extent_page_data {
 120         struct bio *bio;
 121         struct extent_io_tree *tree;
 122         get_extent_t *get_extent;
 123         unsigned long bio_flags;
 124
 125         /* tells writepage not to lock the state bits for this range
 126          * it still does the unlocking
 127          */
 128         unsigned int extent_locked:1;
 129
 130         /* tells the submit_bio code to use a WRITE_SYNC */
 131         unsigned int sync_io:1;
 132 };
 133
 134 static void add_extent_changeset(struct extent_state *state, unsigned bits,
 135                                  struct extent_changeset *changeset,
 136                                  int set)
 137 {
 138         int ret;
 139
 140         if (!changeset)
 141                 return;
 142         if (set && (state->state & bits) == bits)
 143                 return;
 144         if (!set && (state->state & bits) == 0)
 145                 return;
 146         changeset->bytes_changed += state->end - state->start + 1;
 147         ret = ulist_add(changeset->range_changed, state->start, state->end,
 148                         GFP_ATOMIC);
 149         /* ENOMEM */
 150         BUG_ON(ret < 0);
 151 }
 152
 153 static noinline void flush_write_bio(void *data);
 154 static inline struct btrfs_fs_info *
 155 tree_fs_info(struct extent_io_tree *tree)
 156 {
 157         if (!tree->mapping)
 158                 return NULL;
 159         return btrfs_sb(tree->mapping->host->i_sb);
 160 }
 161
 162 int __init extent_io_init(void)
 163 {
 164         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 165                         sizeof(struct extent_state), 0,
 166                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 167         if (!extent_state_cache)
 168                 return -ENOMEM;
 169
 170         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 171                         sizeof(struct extent_buffer), 0,
 172                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 173         if (!extent_buffer_cache)
 174                 goto free_state_cache;
 175
 176         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 177                                      offsetof(struct btrfs_io_bio, bio));
 178         if (!btrfs_bioset)
 179                 goto free_buffer_cache;
 180
 181         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 182                 goto free_bioset;
 183
 184         return 0;
 185
 186 free_bioset:
 187         bioset_free(btrfs_bioset);
 188         btrfs_bioset = NULL;
 189
 190 free_buffer_cache:
 191         kmem_cache_destroy(extent_buffer_cache);
 192         extent_buffer_cache = NULL;
 193
 194 free_state_cache:
 195         kmem_cache_destroy(extent_state_cache);
 196         extent_state_cache = NULL;
 197         return -ENOMEM;
 198 }
 199
 200 void extent_io_exit(void)
 201 {
 202         btrfs_leak_debug_check();
 203
 204         /*
 205          * Make sure all delayed rcu free are flushed before we
 206          * destroy caches.
 207          */
 208         rcu_barrier();
 209         if (extent_state_cache)
 210                 kmem_cache_destroy(extent_state_cache);
 211         if (extent_buffer_cache)
 212                 kmem_cache_destroy(extent_buffer_cache);
 213         if (btrfs_bioset)
 214                 bioset_free(btrfs_bioset);
 215 }
 216
 217 void extent_io_tree_init(struct extent_io_tree *tree,
 218                          struct address_space *mapping)
 219 {
 220         tree->state = RB_ROOT;
 221         tree->ops = NULL;
 222         tree->dirty_bytes = 0;
 223         spin_lock_init(&tree->lock);
 224         tree->mapping = mapping;
 225 }
 226
 227 static struct extent_state *alloc_extent_state(gfp_t mask)
 228 {
 229         struct extent_state *state;
 230
 231         state = kmem_cache_alloc(extent_state_cache, mask);
 232         if (!state)
 233                 return state;
 234         state->state = 0;
 235         state->private = 0;
 236         RB_CLEAR_NODE(&state->rb_node);
 237         btrfs_leak_debug_add(&state->leak_list, &states);
 238         atomic_set(&state->refs, 1);
 239         init_waitqueue_head(&state->wq);
 240         trace_alloc_extent_state(state, mask, _RET_IP_);
 241         return state;
 242 }
 243
 244 void free_extent_state(struct extent_state *state)
 245 {
 246         if (!state)
 247                 return;
 248         if (atomic_dec_and_test(&state->refs)) {
 249                 WARN_ON(extent_state_in_tree(state));
 250                 btrfs_leak_debug_del(&state->leak_list);
 251                 trace_free_extent_state(state, _RET_IP_);
 252                 kmem_cache_free(extent_state_cache, state);
 253         }
 254 }
 255
 256 static struct rb_node *tree_insert(struct rb_root *root,
 257                                    struct rb_node *search_start,
 258                                    u64 offset,
 259                                    struct rb_node *node,
 260                                    struct rb_node ***p_in,
 261                                    struct rb_node **parent_in)
 262 {
 263         struct rb_node **p;
 264         struct rb_node *parent = NULL;
 265         struct tree_entry *entry;
 266
 267         if (p_in && parent_in) {
 268                 p = *p_in;
 269                 parent = *parent_in;
 270                 goto do_insert;
 271         }
 272
 273         p = search_start ? &search_start : &root->rb_node;
 274         while (*p) {
 275                 parent = *p;
 276                 entry = rb_entry(parent, struct tree_entry, rb_node);
 277
 278                 if (offset < entry->start)
 279                         p = &(*p)->rb_left;
 280                 else if (offset > entry->end)
 281                         p = &(*p)->rb_right;
 282                 else
 283                         return parent;
 284         }
 285
 286 do_insert:
 287         rb_link_node(node, parent, p);
 288         rb_insert_color(node, root);
 289         return NULL;
 290 }
 291
 292 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 293                                       struct rb_node **prev_ret,
 294                                       struct rb_node **next_ret,
 295                                       struct rb_node ***p_ret,
 296                                       struct rb_node **parent_ret)
 297 {
 298         struct rb_root *root = &tree->state;
 299         struct rb_node **n = &root->rb_node;
 300         struct rb_node *prev = NULL;
 301         struct rb_node *orig_prev = NULL;
 302         struct tree_entry *entry;
 303         struct tree_entry *prev_entry = NULL;
 304
 305         while (*n) {
 306                 prev = *n;
 307                 entry = rb_entry(prev, struct tree_entry, rb_node);
 308                 prev_entry = entry;
 309
 310                 if (offset < entry->start)
 311                         n = &(*n)->rb_left;
 312                 else if (offset > entry->end)
 313                         n = &(*n)->rb_right;
 314                 else
 315                         return *n;
 316         }
 317
 318         if (p_ret)
 319                 *p_ret = n;
 320         if (parent_ret)
 321                 *parent_ret = prev;
 322
 323         if (prev_ret) {
 324                 orig_prev = prev;
 325                 while (prev && offset > prev_entry->end) {
 326                         prev = rb_next(prev);
 327                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 328                 }
 329                 *prev_ret = prev;
 330                 prev = orig_prev;
 331         }
 332
 333         if (next_ret) {
 334                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 335                 while (prev && offset < prev_entry->start) {
 336                         prev = rb_prev(prev);
 337                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 338                 }
 339                 *next_ret = prev;
 340         }
 341         return NULL;
 342 }
 343
 344 static inline struct rb_node *
 345 tree_search_for_insert(struct extent_io_tree *tree,
 346                        u64 offset,
 347                        struct rb_node ***p_ret,
 348                        struct rb_node **parent_ret)
 349 {
 350         struct rb_node *prev = NULL;
 351         struct rb_node *ret;
 352
 353         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 354         if (!ret)
 355                 return prev;
 356         return ret;
 357 }
 358
 359 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 360                                           u64 offset)
 361 {
 362         return tree_search_for_insert(tree, offset, NULL, NULL);
 363 }
 364
 365 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 366                      struct extent_state *other)
 367 {
 368         if (tree->ops && tree->ops->merge_extent_hook)
 369                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 370                                              other);
 371 }
 372
 373 /*
 374  * utility function to look for merge candidates inside a given range.
 375  * Any extents with matching state are merged together into a single
 376  * extent in the tree.  Extents with EXTENT_IO in their state field
 377  * are not merged because the end_io handlers need to be able to do
 378  * operations on them without sleeping (or doing allocations/splits).
 379  *
 380  * This should be called with the tree lock held.
 381  */
 382 static void merge_state(struct extent_io_tree *tree,
 383                         struct extent_state *state)
 384 {
 385         struct extent_state *other;
 386         struct rb_node *other_node;
 387
 388         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 389                 return;
 390
 391         other_node = rb_prev(&state->rb_node);
 392         if (other_node) {
 393                 other = rb_entry(other_node, struct extent_state, rb_node);
 394                 if (other->end == state->start - 1 &&
 395                     other->state == state->state) {
 396                         merge_cb(tree, state, other);
 397                         state->start = other->start;
 398                         rb_erase(&other->rb_node, &tree->state);
 399                         RB_CLEAR_NODE(&other->rb_node);
 400                         free_extent_state(other);
 401                 }
 402         }
 403         other_node = rb_next(&state->rb_node);
 404         if (other_node) {
 405                 other = rb_entry(other_node, struct extent_state, rb_node);
 406                 if (other->start == state->end + 1 &&
 407                     other->state == state->state) {
 408                         merge_cb(tree, state, other);
 409                         state->end = other->end;
 410                         rb_erase(&other->rb_node, &tree->state);
 411                         RB_CLEAR_NODE(&other->rb_node);
 412                         free_extent_state(other);
 413                 }
 414         }
 415 }
 416
 417 static void set_state_cb(struct extent_io_tree *tree,
 418                          struct extent_state *state, unsigned *bits)
 419 {
 420         if (tree->ops && tree->ops->set_bit_hook)
 421                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 422 }
 423
 424 static void clear_state_cb(struct extent_io_tree *tree,
 425                            struct extent_state *state, unsigned *bits)
 426 {
 427         if (tree->ops && tree->ops->clear_bit_hook)
 428                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 429 }
 430
 431 static void set_state_bits(struct extent_io_tree *tree,
 432                            struct extent_state *state, unsigned *bits,
 433                            struct extent_changeset *changeset);
 434
 435 /*
 436  * insert an extent_state struct into the tree.  'bits' are set on the
 437  * struct before it is inserted.
 438  *
 439  * This may return -EEXIST if the extent is already there, in which case the
 440  * state struct is freed.
 441  *
 442  * The tree lock is not taken internally.  This is a utility function and
 443  * probably isn't what you want to call (see set/clear_extent_bit).
 444  */
 445 static int insert_state(struct extent_io_tree *tree,
 446                         struct extent_state *state, u64 start, u64 end,
 447                         struct rb_node ***p,
 448                         struct rb_node **parent,
 449                         unsigned *bits, struct extent_changeset *changeset)
 450 {
 451         struct rb_node *node;
 452
 453         if (end < start)
 454                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 455                        end, start);
 456         state->start = start;
 457         state->end = end;
 458
 459         set_state_bits(tree, state, bits, changeset);
 460
 461         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 462         if (node) {
 463                 struct extent_state *found;
 464                 found = rb_entry(node, struct extent_state, rb_node);
 465                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
 466                        "%llu %llu\n",
 467                        found->start, found->end, start, end);
 468                 return -EEXIST;
 469         }
 470         merge_state(tree, state);
 471         return 0;
 472 }
 473
 474 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 475                      u64 split)
 476 {
 477         if (tree->ops && tree->ops->split_extent_hook)
 478                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 479 }
 480
 481 /*
 482  * split a given extent state struct in two, inserting the preallocated
 483  * struct 'prealloc' as the newly created second half.  'split' indicates an
 484  * offset inside 'orig' where it should be split.
 485  *
 486  * Before calling,
 487  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 488  * are two extent state structs in the tree:
 489  * prealloc: [orig->start, split - 1]
 490  * orig: [ split, orig->end ]
 491  *
 492  * The tree locks are not taken by this function. They need to be held
 493  * by the caller.
 494  */
 495 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 496                        struct extent_state *prealloc, u64 split)
 497 {
 498         struct rb_node *node;
 499
 500         split_cb(tree, orig, split);
 501
 502         prealloc->start = orig->start;
 503         prealloc->end = split - 1;
 504         prealloc->state = orig->state;
 505         orig->start = split;
 506
 507         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 508                            &prealloc->rb_node, NULL, NULL);
 509         if (node) {
 510                 free_extent_state(prealloc);
 511                 return -EEXIST;
 512         }
 513         return 0;
 514 }
 515
 516 static struct extent_state *next_state(struct extent_state *state)
 517 {
 518         struct rb_node *next = rb_next(&state->rb_node);
 519         if (next)
 520                 return rb_entry(next, struct extent_state, rb_node);
 521         else
 522                 return NULL;
 523 }
 524
 525 /*
 526  * utility function to clear some bits in an extent state struct.
 527  * it will optionally wake up any one waiting on this state (wake == 1).
 528  *
 529  * If no bits are set on the state struct after clearing things, the
 530  * struct is freed and removed from the tree
 531  */
 532 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 533                                             struct extent_state *state,
 534                                             unsigned *bits, int wake,
 535                                             struct extent_changeset *changeset)
 536 {
 537         struct extent_state *next;
 538         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 539
 540         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 541                 u64 range = state->end - state->start + 1;
 542                 WARN_ON(range > tree->dirty_bytes);
 543                 tree->dirty_bytes -= range;
 544         }
 545         clear_state_cb(tree, state, bits);
 546         add_extent_changeset(state, bits_to_clear, changeset, 0);
 547         state->state &= ~bits_to_clear;
 548         if (wake)
 549                 wake_up(&state->wq);
 550         if (state->state == 0) {
 551                 next = next_state(state);
 552                 if (extent_state_in_tree(state)) {
 553                         rb_erase(&state->rb_node, &tree->state);
 554                         RB_CLEAR_NODE(&state->rb_node);
 555                         free_extent_state(state);
 556                 } else {
 557                         WARN_ON(1);
 558                 }
 559         } else {
 560                 merge_state(tree, state);
 561                 next = next_state(state);
 562         }
 563         return next;
 564 }
 565
 566 static struct extent_state *
 567 alloc_extent_state_atomic(struct extent_state *prealloc)
 568 {
 569         if (!prealloc)
 570                 prealloc = alloc_extent_state(GFP_ATOMIC);
 571
 572         return prealloc;
 573 }
 574
 575 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 576 {
 577         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 578                     "Extent tree was modified by another "
 579                     "thread while locked.");
 580 }
 581
 582 /*
 583  * clear some bits on a range in the tree.  This may require splitting
 584  * or inserting elements in the tree, so the gfp mask is used to
 585  * indicate which allocations or sleeping are allowed.
 586  *
 587  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 588  * the given range from the tree regardless of state (ie for truncate).
 589  *
 590  * the range [start, end] is inclusive.
 591  *
 592  * This takes the tree lock, and returns 0 on success and < 0 on error.
 593  */
 594 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 595                               unsigned bits, int wake, int delete,
 596                               struct extent_state **cached_state,
 597                               gfp_t mask, struct extent_changeset *changeset)
 598 {
 599         struct extent_state *state;
 600         struct extent_state *cached;
 601         struct extent_state *prealloc = NULL;
 602         struct rb_node *node;
 603         u64 last_end;
 604         int err;
 605         int clear = 0;
 606
 607         btrfs_debug_check_extent_io_range(tree, start, end);
 608
 609         if (bits & EXTENT_DELALLOC)
 610                 bits |= EXTENT_NORESERVE;
 611
 612         if (delete)
 613                 bits |= ~EXTENT_CTLBITS;
 614         bits |= EXTENT_FIRST_DELALLOC;
 615
 616         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 617                 clear = 1;
 618 again:
 619         if (!prealloc && gfpflags_allow_blocking(mask)) {
 620                 /*
 621                  * Don't care for allocation failure here because we might end
 622                  * up not needing the pre-allocated extent state at all, which
 623                  * is the case if we only have in the tree extent states that
 624                  * cover our input range and don't cover too any other range.
 625                  * If we end up needing a new extent state we allocate it later.
 626                  */
 627                 prealloc = alloc_extent_state(mask);
 628         }
 629
 630         spin_lock(&tree->lock);
 631         if (cached_state) {
 632                 cached = *cached_state;
 633
 634                 if (clear) {
 635                         *cached_state = NULL;
 636                         cached_state = NULL;
 637                 }
 638
 639                 if (cached && extent_state_in_tree(cached) &&
 640                     cached->start <= start && cached->end > start) {
 641                         if (clear)
 642                                 atomic_dec(&cached->refs);
 643                         state = cached;
 644                         goto hit_next;
 645                 }
 646                 if (clear)
 647                         free_extent_state(cached);
 648         }
 649         /*
 650          * this search will find the extents that end after
 651          * our range starts
 652          */
 653         node = tree_search(tree, start);
 654         if (!node)
 655                 goto out;
 656         state = rb_entry(node, struct extent_state, rb_node);
 657 hit_next:
 658         if (state->start > end)
 659                 goto out;
 660         WARN_ON(state->end < start);
 661         last_end = state->end;
 662
 663         /* the state doesn't have the wanted bits, go ahead */
 664         if (!(state->state & bits)) {
 665                 state = next_state(state);
 666                 goto next;
 667         }
 668
 669         /*
 670          *     | ---- desired range ---- |
 671          *  | state | or
 672          *  | ------------- state -------------- |
 673          *
 674          * We need to split the extent we found, and may flip
 675          * bits on second half.
 676          *
 677          * If the extent we found extends past our range, we
 678          * just split and search again.  It'll get split again
 679          * the next time though.
 680          *
 681          * If the extent we found is inside our range, we clear
 682          * the desired bit on it.
 683          */
 684
 685         if (state->start < start) {
 686                 prealloc = alloc_extent_state_atomic(prealloc);
 687                 BUG_ON(!prealloc);
 688                 err = split_state(tree, state, prealloc, start);
 689                 if (err)
 690                         extent_io_tree_panic(tree, err);
 691
 692                 prealloc = NULL;
 693                 if (err)
 694                         goto out;
 695                 if (state->end <= end) {
 696                         state = clear_state_bit(tree, state, &bits, wake,
 697                                                 changeset);
 698                         goto next;
 699                 }
 700                 goto search_again;
 701         }
 702         /*
 703          * | ---- desired range ---- |
 704          *                        | state |
 705          * We need to split the extent, and clear the bit
 706          * on the first half
 707          */
 708         if (state->start <= end && state->end > end) {
 709                 prealloc = alloc_extent_state_atomic(prealloc);
 710                 BUG_ON(!prealloc);
 711                 err = split_state(tree, state, prealloc, end + 1);
 712                 if (err)
 713                         extent_io_tree_panic(tree, err);
 714
 715                 if (wake)
 716                         wake_up(&state->wq);
 717
 718                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 719
 720                 prealloc = NULL;
 721                 goto out;
 722         }
 723
 724         state = clear_state_bit(tree, state, &bits, wake, changeset);
 725 next:
 726         if (last_end == (u64)-1)
 727                 goto out;
 728         start = last_end + 1;
 729         if (start <= end && state && !need_resched())
 730                 goto hit_next;
 731         goto search_again;
 732
 733 out:
 734         spin_unlock(&tree->lock);
 735         if (prealloc)
 736                 free_extent_state(prealloc);
 737
 738         return 0;
 739
 740 search_again:
 741         if (start > end)
 742                 goto out;
 743         spin_unlock(&tree->lock);
 744         if (gfpflags_allow_blocking(mask))
 745                 cond_resched();
 746         goto again;
 747 }
 748
 749 static void wait_on_state(struct extent_io_tree *tree,
 750                           struct extent_state *state)
 751                 __releases(tree->lock)
 752                 __acquires(tree->lock)
 753 {
 754         DEFINE_WAIT(wait);
 755         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 756         spin_unlock(&tree->lock);
 757         schedule();
 758         spin_lock(&tree->lock);
 759         finish_wait(&state->wq, &wait);
 760 }
 761
 762 /*
 763  * waits for one or more bits to clear on a range in the state tree.
 764  * The range [start, end] is inclusive.
 765  * The tree lock is taken by this function
 766  */
 767 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 768                             unsigned long bits)
 769 {
 770         struct extent_state *state;
 771         struct rb_node *node;
 772
 773         btrfs_debug_check_extent_io_range(tree, start, end);
 774
 775         spin_lock(&tree->lock);
 776 again:
 777         while (1) {
 778                 /*
 779                  * this search will find all the extents that end after
 780                  * our range starts
 781                  */
 782                 node = tree_search(tree, start);
 783 process_node:
 784                 if (!node)
 785                         break;
 786
 787                 state = rb_entry(node, struct extent_state, rb_node);
 788
 789                 if (state->start > end)
 790                         goto out;
 791
 792                 if (state->state & bits) {
 793                         start = state->start;
 794                         atomic_inc(&state->refs);
 795                         wait_on_state(tree, state);
 796                         free_extent_state(state);
 797                         goto again;
 798                 }
 799                 start = state->end + 1;
 800
 801                 if (start > end)
 802                         break;
 803
 804                 if (!cond_resched_lock(&tree->lock)) {
 805                         node = rb_next(node);
 806                         goto process_node;
 807                 }
 808         }
 809 out:
 810         spin_unlock(&tree->lock);
 811 }
 812
 813 static void set_state_bits(struct extent_io_tree *tree,
 814                            struct extent_state *state,
 815                            unsigned *bits, struct extent_changeset *changeset)
 816 {
 817         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 818
 819         set_state_cb(tree, state, bits);
 820         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 821                 u64 range = state->end - state->start + 1;
 822                 tree->dirty_bytes += range;
 823         }
 824         add_extent_changeset(state, bits_to_set, changeset, 1);
 825         state->state |= bits_to_set;
 826 }
 827
 828 static void cache_state_if_flags(struct extent_state *state,
 829                                  struct extent_state **cached_ptr,
 830                                  unsigned flags)
 831 {
 832         if (cached_ptr && !(*cached_ptr)) {
 833                 if (!flags || (state->state & flags)) {
 834                         *cached_ptr = state;
 835                         atomic_inc(&state->refs);
 836                 }
 837         }
 838 }
 839
 840 static void cache_state(struct extent_state *state,
 841                         struct extent_state **cached_ptr)
 842 {
 843         return cache_state_if_flags(state, cached_ptr,
 844                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 845 }
 846
 847 /*
 848  * set some bits on a range in the tree.  This may require allocations or
 849  * sleeping, so the gfp mask is used to indicate what is allowed.
 850  *
 851  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 852  * part of the range already has the desired bits set.  The start of the
 853  * existing range is returned in failed_start in this case.
 854  *
 855  * [start, end] is inclusive This takes the tree lock.
 856  */
 857
 858 static int __must_check
 859 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 860                  unsigned bits, unsigned exclusive_bits,
 861                  u64 *failed_start, struct extent_state **cached_state,
 862                  gfp_t mask, struct extent_changeset *changeset)
 863 {
 864         struct extent_state *state;
 865         struct extent_state *prealloc = NULL;
 866         struct rb_node *node;
 867         struct rb_node **p;
 868         struct rb_node *parent;
 869         int err = 0;
 870         u64 last_start;
 871         u64 last_end;
 872
 873         btrfs_debug_check_extent_io_range(tree, start, end);
 874
 875         bits |= EXTENT_FIRST_DELALLOC;
 876 again:
 877         if (!prealloc && gfpflags_allow_blocking(mask)) {
 878                 prealloc = alloc_extent_state(mask);
 879                 BUG_ON(!prealloc);
 880         }
 881
 882         spin_lock(&tree->lock);
 883         if (cached_state && *cached_state) {
 884                 state = *cached_state;
 885                 if (state->start <= start && state->end > start &&
 886                     extent_state_in_tree(state)) {
 887                         node = &state->rb_node;
 888                         goto hit_next;
 889                 }
 890         }
 891         /*
 892          * this search will find all the extents that end after
 893          * our range starts.
 894          */
 895         node = tree_search_for_insert(tree, start, &p, &parent);
 896         if (!node) {
 897                 prealloc = alloc_extent_state_atomic(prealloc);
 898                 BUG_ON(!prealloc);
 899                 err = insert_state(tree, prealloc, start, end,
 900                                    &p, &parent, &bits, changeset);
 901                 if (err)
 902                         extent_io_tree_panic(tree, err);
 903
 904                 cache_state(prealloc, cached_state);
 905                 prealloc = NULL;
 906                 goto out;
 907         }
 908         state = rb_entry(node, struct extent_state, rb_node);
 909 hit_next:
 910         last_start = state->start;
 911         last_end = state->end;
 912
 913         /*
 914          * | ---- desired range ---- |
 915          * | state |
 916          *
 917          * Just lock what we found and keep going
 918          */
 919         if (state->start == start && state->end <= end) {
 920                 if (state->state & exclusive_bits) {
 921                         *failed_start = state->start;
 922                         err = -EEXIST;
 923                         goto out;
 924                 }
 925
 926                 set_state_bits(tree, state, &bits, changeset);
 927                 cache_state(state, cached_state);
 928                 merge_state(tree, state);
 929                 if (last_end == (u64)-1)
 930                         goto out;
 931                 start = last_end + 1;
 932                 state = next_state(state);
 933                 if (start < end && state && state->start == start &&
 934                     !need_resched())
 935                         goto hit_next;
 936                 goto search_again;
 937         }
 938
 939         /*
 940          *     | ---- desired range ---- |
 941          * | state |
 942          *   or
 943          * | ------------- state -------------- |
 944          *
 945          * We need to split the extent we found, and may flip bits on
 946          * second half.
 947          *
 948          * If the extent we found extends past our
 949          * range, we just split and search again.  It'll get split
 950          * again the next time though.
 951          *
 952          * If the extent we found is inside our range, we set the
 953          * desired bit on it.
 954          */
 955         if (state->start < start) {
 956                 if (state->state & exclusive_bits) {
 957                         *failed_start = start;
 958                         err = -EEXIST;
 959                         goto out;
 960                 }
 961
 962                 prealloc = alloc_extent_state_atomic(prealloc);
 963                 BUG_ON(!prealloc);
 964                 err = split_state(tree, state, prealloc, start);
 965                 if (err)
 966                         extent_io_tree_panic(tree, err);
 967
 968                 prealloc = NULL;
 969                 if (err)
 970                         goto out;
 971                 if (state->end <= end) {
 972                         set_state_bits(tree, state, &bits, changeset);
 973                         cache_state(state, cached_state);
 974                         merge_state(tree, state);
 975                         if (last_end == (u64)-1)
 976                                 goto out;
 977                         start = last_end + 1;
 978                         state = next_state(state);
 979                         if (start < end && state && state->start == start &&
 980                             !need_resched())
 981                                 goto hit_next;
 982                 }
 983                 goto search_again;
 984         }
 985         /*
 986          * | ---- desired range ---- |
 987          *     | state | or               | state |
 988          *
 989          * There's a hole, we need to insert something in it and
 990          * ignore the extent we found.
 991          */
 992         if (state->start > start) {
 993                 u64 this_end;
 994                 if (end < last_start)
 995                         this_end = end;
 996                 else
 997                         this_end = last_start - 1;
 998
 999                 prealloc = alloc_extent_state_atomic(prealloc);
1000                 BUG_ON(!prealloc);
1001
1002                 /*
1003                  * Avoid to free 'prealloc' if it can be merged with
1004                  * the later extent.
1005                  */
1006                 err = insert_state(tree, prealloc, start, this_end,
1007                                    NULL, NULL, &bits, changeset);
1008                 if (err)
1009                         extent_io_tree_panic(tree, err);
1010
1011                 cache_state(prealloc, cached_state);
1012                 prealloc = NULL;
1013                 start = this_end + 1;
1014                 goto search_again;
1015         }
1016         /*
1017          * | ---- desired range ---- |
1018          *                        | state |
1019          * We need to split the extent, and set the bit
1020          * on the first half
1021          */
1022         if (state->start <= end && state->end > end) {
1023                 if (state->state & exclusive_bits) {
1024                         *failed_start = start;
1025                         err = -EEXIST;
1026                         goto out;
1027                 }
1028
1029                 prealloc = alloc_extent_state_atomic(prealloc);
1030                 BUG_ON(!prealloc);
1031                 err = split_state(tree, state, prealloc, end + 1);
1032                 if (err)
1033                         extent_io_tree_panic(tree, err);
1034
1035                 set_state_bits(tree, prealloc, &bits, changeset);
1036                 cache_state(prealloc, cached_state);
1037                 merge_state(tree, prealloc);
1038                 prealloc = NULL;
1039                 goto out;
1040         }
1041
1042         goto search_again;
1043
1044 out:
1045         spin_unlock(&tree->lock);
1046         if (prealloc)
1047                 free_extent_state(prealloc);
1048
1049         return err;
1050
1051 search_again:
1052         if (start > end)
1053                 goto out;
1054         spin_unlock(&tree->lock);
1055         if (gfpflags_allow_blocking(mask))
1056                 cond_resched();
1057         goto again;
1058 }
1059
1060 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1061                    unsigned bits, u64 * failed_start,
1062                    struct extent_state **cached_state, gfp_t mask)
1063 {
1064         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1065                                 cached_state, mask, NULL);
1066 }
1067
1068
1069 /**
1070  * convert_extent_bit - convert all bits in a given range from one bit to
1071  *                      another
1072  * @tree:       the io tree to search
1073  * @start:      the start offset in bytes
1074  * @end:        the end offset in bytes (inclusive)
1075  * @bits:       the bits to set in this range
1076  * @clear_bits: the bits to clear in this range
1077  * @cached_state:       state that we're going to cache
1078  * @mask:       the allocation mask
1079  *
1080  * This will go through and set bits for the given range.  If any states exist
1081  * already in this range they are set with the given bit and cleared of the
1082  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1083  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1084  * boundary bits like LOCK.
1085  */
1086 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1087                        unsigned bits, unsigned clear_bits,
1088                        struct extent_state **cached_state, gfp_t mask)
1089 {
1090         struct extent_state *state;
1091         struct extent_state *prealloc = NULL;
1092         struct rb_node *node;
1093         struct rb_node **p;
1094         struct rb_node *parent;
1095         int err = 0;
1096         u64 last_start;
1097         u64 last_end;
1098         bool first_iteration = true;
1099
1100         btrfs_debug_check_extent_io_range(tree, start, end);
1101
1102 again:
1103         if (!prealloc && gfpflags_allow_blocking(mask)) {
1104                 /*
1105                  * Best effort, don't worry if extent state allocation fails
1106                  * here for the first iteration. We might have a cached state
1107                  * that matches exactly the target range, in which case no
1108                  * extent state allocations are needed. We'll only know this
1109                  * after locking the tree.
1110                  */
1111                 prealloc = alloc_extent_state(mask);
1112                 if (!prealloc && !first_iteration)
1113                         return -ENOMEM;
1114         }
1115
1116         spin_lock(&tree->lock);
1117         if (cached_state && *cached_state) {
1118                 state = *cached_state;
1119                 if (state->start <= start && state->end > start &&
1120                     extent_state_in_tree(state)) {
1121                         node = &state->rb_node;
1122                         goto hit_next;
1123                 }
1124         }
1125
1126         /*
1127          * this search will find all the extents that end after
1128          * our range starts.
1129          */
1130         node = tree_search_for_insert(tree, start, &p, &parent);
1131         if (!node) {
1132                 prealloc = alloc_extent_state_atomic(prealloc);
1133                 if (!prealloc) {
1134                         err = -ENOMEM;
1135                         goto out;
1136                 }
1137                 err = insert_state(tree, prealloc, start, end,
1138                                    &p, &parent, &bits, NULL);
1139                 if (err)
1140                         extent_io_tree_panic(tree, err);
1141                 cache_state(prealloc, cached_state);
1142                 prealloc = NULL;
1143                 goto out;
1144         }
1145         state = rb_entry(node, struct extent_state, rb_node);
1146 hit_next:
1147         last_start = state->start;
1148         last_end = state->end;
1149
1150         /*
1151          * | ---- desired range ---- |
1152          * | state |
1153          *
1154          * Just lock what we found and keep going
1155          */
1156         if (state->start == start && state->end <= end) {
1157                 set_state_bits(tree, state, &bits, NULL);
1158                 cache_state(state, cached_state);
1159                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1160                 if (last_end == (u64)-1)
1161                         goto out;
1162                 start = last_end + 1;
1163                 if (start < end && state && state->start == start &&
1164                     !need_resched())
1165                         goto hit_next;
1166                 goto search_again;
1167         }
1168
1169         /*
1170          *     | ---- desired range ---- |
1171          * | state |
1172          *   or
1173          * | ------------- state -------------- |
1174          *
1175          * We need to split the extent we found, and may flip bits on
1176          * second half.
1177          *
1178          * If the extent we found extends past our
1179          * range, we just split and search again.  It'll get split
1180          * again the next time though.
1181          *
1182          * If the extent we found is inside our range, we set the
1183          * desired bit on it.
1184          */
1185         if (state->start < start) {
1186                 prealloc = alloc_extent_state_atomic(prealloc);
1187                 if (!prealloc) {
1188                         err = -ENOMEM;
1189                         goto out;
1190                 }
1191                 err = split_state(tree, state, prealloc, start);
1192                 if (err)
1193                         extent_io_tree_panic(tree, err);
1194                 prealloc = NULL;
1195                 if (err)
1196                         goto out;
1197                 if (state->end <= end) {
1198                         set_state_bits(tree, state, &bits, NULL);
1199                         cache_state(state, cached_state);
1200                         state = clear_state_bit(tree, state, &clear_bits, 0,
1201                                                 NULL);
1202                         if (last_end == (u64)-1)
1203                                 goto out;
1204                         start = last_end + 1;
1205                         if (start < end && state && state->start == start &&
1206                             !need_resched())
1207                                 goto hit_next;
1208                 }
1209                 goto search_again;
1210         }
1211         /*
1212          * | ---- desired range ---- |
1213          *     | state | or               | state |
1214          *
1215          * There's a hole, we need to insert something in it and
1216          * ignore the extent we found.
1217          */
1218         if (state->start > start) {
1219                 u64 this_end;
1220                 if (end < last_start)
1221                         this_end = end;
1222                 else
1223                         this_end = last_start - 1;
1224
1225                 prealloc = alloc_extent_state_atomic(prealloc);
1226                 if (!prealloc) {
1227                         err = -ENOMEM;
1228                         goto out;
1229                 }
1230
1231                 /*
1232                  * Avoid to free 'prealloc' if it can be merged with
1233                  * the later extent.
1234                  */
1235                 err = insert_state(tree, prealloc, start, this_end,
1236                                    NULL, NULL, &bits, NULL);
1237                 if (err)
1238                         extent_io_tree_panic(tree, err);
1239                 cache_state(prealloc, cached_state);
1240                 prealloc = NULL;
1241                 start = this_end + 1;
1242                 goto search_again;
1243         }
1244         /*
1245          * | ---- desired range ---- |
1246          *                        | state |
1247          * We need to split the extent, and set the bit
1248          * on the first half
1249          */
1250         if (state->start <= end && state->end > end) {
1251                 prealloc = alloc_extent_state_atomic(prealloc);
1252                 if (!prealloc) {
1253                         err = -ENOMEM;
1254                         goto out;
1255                 }
1256
1257                 err = split_state(tree, state, prealloc, end + 1);
1258                 if (err)
1259                         extent_io_tree_panic(tree, err);
1260
1261                 set_state_bits(tree, prealloc, &bits, NULL);
1262                 cache_state(prealloc, cached_state);
1263                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1264                 prealloc = NULL;
1265                 goto out;
1266         }
1267
1268         goto search_again;
1269
1270 out:
1271         spin_unlock(&tree->lock);
1272         if (prealloc)
1273                 free_extent_state(prealloc);
1274
1275         return err;
1276
1277 search_again:
1278         if (start > end)
1279                 goto out;
1280         spin_unlock(&tree->lock);
1281         if (gfpflags_allow_blocking(mask))
1282                 cond_resched();
1283         first_iteration = false;
1284         goto again;
1285 }
1286
1287 /* wrappers around set/clear extent bit */
1288 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1289                            unsigned bits, gfp_t mask,
1290                            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, mask,
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, gfp_t mask,
1314                              struct extent_changeset *changeset)
1315 {
1316         /*
1317          * Don't support EXTENT_LOCKED case, same reason as
1318          * set_record_extent_bits().
1319          */
1320         BUG_ON(bits & EXTENT_LOCKED);
1321
1322         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask,
1323                                   changeset);
1324 }
1325
1326 /*
1327  * either insert or lock state struct between start and end use mask to tell
1328  * us if waiting is desired.
1329  */
1330 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1331                      struct extent_state **cached_state)
1332 {
1333         int err;
1334         u64 failed_start;
1335
1336         while (1) {
1337                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1338                                        EXTENT_LOCKED, &failed_start,
1339                                        cached_state, GFP_NOFS, NULL);
1340                 if (err == -EEXIST) {
1341                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1342                         start = failed_start;
1343                 } else
1344                         break;
1345                 WARN_ON(start > end);
1346         }
1347         return err;
1348 }
1349
1350 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1351 {
1352         int err;
1353         u64 failed_start;
1354
1355         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1356                                &failed_start, NULL, GFP_NOFS, NULL);
1357         if (err == -EEXIST) {
1358                 if (failed_start > start)
1359                         clear_extent_bit(tree, start, failed_start - 1,
1360                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1361                 return 0;
1362         }
1363         return 1;
1364 }
1365
1366 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1367 {
1368         unsigned long index = start >> PAGE_CACHE_SHIFT;
1369         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1370         struct page *page;
1371
1372         while (index <= end_index) {
1373                 page = find_get_page(inode->i_mapping, index);
1374                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1375                 clear_page_dirty_for_io(page);
1376                 page_cache_release(page);
1377                 index++;
1378         }
1379 }
1380
1381 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1382 {
1383         unsigned long index = start >> PAGE_CACHE_SHIFT;
1384         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1385         struct page *page;
1386
1387         while (index <= end_index) {
1388                 page = find_get_page(inode->i_mapping, index);
1389                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1390                 __set_page_dirty_nobuffers(page);
1391                 account_page_redirty(page);
1392                 page_cache_release(page);
1393                 index++;
1394         }
1395 }
1396
1397 /*
1398  * helper function to set both pages and extents in the tree writeback
1399  */
1400 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1401 {
1402         unsigned long index = start >> PAGE_CACHE_SHIFT;
1403         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1404         struct page *page;
1405
1406         while (index <= end_index) {
1407                 page = find_get_page(tree->mapping, index);
1408                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1409                 set_page_writeback(page);
1410                 page_cache_release(page);
1411                 index++;
1412         }
1413 }
1414
1415 /* find the first state struct with 'bits' set after 'start', and
1416  * return it.  tree->lock must be held.  NULL will returned if
1417  * nothing was found after 'start'
1418  */
1419 static struct extent_state *
1420 find_first_extent_bit_state(struct extent_io_tree *tree,
1421                             u64 start, unsigned bits)
1422 {
1423         struct rb_node *node;
1424         struct extent_state *state;
1425
1426         /*
1427          * this search will find all the extents that end after
1428          * our range starts.
1429          */
1430         node = tree_search(tree, start);
1431         if (!node)
1432                 goto out;
1433
1434         while (1) {
1435                 state = rb_entry(node, struct extent_state, rb_node);
1436                 if (state->end >= start && (state->state & bits))
1437                         return state;
1438
1439                 node = rb_next(node);
1440                 if (!node)
1441                         break;
1442         }
1443 out:
1444         return NULL;
1445 }
1446
1447 /*
1448  * find the first offset in the io tree with 'bits' set. zero is
1449  * returned if we find something, and *start_ret and *end_ret are
1450  * set to reflect the state struct that was found.
1451  *
1452  * If nothing was found, 1 is returned. If found something, return 0.
1453  */
1454 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1455                           u64 *start_ret, u64 *end_ret, unsigned bits,
1456                           struct extent_state **cached_state)
1457 {
1458         struct extent_state *state;
1459         struct rb_node *n;
1460         int ret = 1;
1461
1462         spin_lock(&tree->lock);
1463         if (cached_state && *cached_state) {
1464                 state = *cached_state;
1465                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1466                         n = rb_next(&state->rb_node);
1467                         while (n) {
1468                                 state = rb_entry(n, struct extent_state,
1469                                                  rb_node);
1470                                 if (state->state & bits)
1471                                         goto got_it;
1472                                 n = rb_next(n);
1473                         }
1474                         free_extent_state(*cached_state);
1475                         *cached_state = NULL;
1476                         goto out;
1477                 }
1478                 free_extent_state(*cached_state);
1479                 *cached_state = NULL;
1480         }
1481
1482         state = find_first_extent_bit_state(tree, start, bits);
1483 got_it:
1484         if (state) {
1485                 cache_state_if_flags(state, cached_state, 0);
1486                 *start_ret = state->start;
1487                 *end_ret = state->end;
1488                 ret = 0;
1489         }
1490 out:
1491         spin_unlock(&tree->lock);
1492         return ret;
1493 }
1494
1495 /*
1496  * find a contiguous range of bytes in the file marked as delalloc, not
1497  * more than 'max_bytes'.  start and end are used to return the range,
1498  *
1499  * 1 is returned if we find something, 0 if nothing was in the tree
1500  */
1501 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1502                                         u64 *start, u64 *end, u64 max_bytes,
1503                                         struct extent_state **cached_state)
1504 {
1505         struct rb_node *node;
1506         struct extent_state *state;
1507         u64 cur_start = *start;
1508         u64 found = 0;
1509         u64 total_bytes = 0;
1510
1511         spin_lock(&tree->lock);
1512
1513         /*
1514          * this search will find all the extents that end after
1515          * our range starts.
1516          */
1517         node = tree_search(tree, cur_start);
1518         if (!node) {
1519                 if (!found)
1520                         *end = (u64)-1;
1521                 goto out;
1522         }
1523
1524         while (1) {
1525                 state = rb_entry(node, struct extent_state, rb_node);
1526                 if (found && (state->start != cur_start ||
1527                               (state->state & EXTENT_BOUNDARY))) {
1528                         goto out;
1529                 }
1530                 if (!(state->state & EXTENT_DELALLOC)) {
1531                         if (!found)
1532                                 *end = state->end;
1533                         goto out;
1534                 }
1535                 if (!found) {
1536                         *start = state->start;
1537                         *cached_state = state;
1538                         atomic_inc(&state->refs);
1539                 }
1540                 found++;
1541                 *end = state->end;
1542                 cur_start = state->end + 1;
1543                 node = rb_next(node);
1544                 total_bytes += state->end - state->start + 1;
1545                 if (total_bytes >= max_bytes)
1546                         break;
1547                 if (!node)
1548                         break;
1549         }
1550 out:
1551         spin_unlock(&tree->lock);
1552         return found;
1553 }
1554
1555 static noinline void __unlock_for_delalloc(struct inode *inode,
1556                                            struct page *locked_page,
1557                                            u64 start, u64 end)
1558 {
1559         int ret;
1560         struct page *pages[16];
1561         unsigned long index = start >> PAGE_CACHE_SHIFT;
1562         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1563         unsigned long nr_pages = end_index - index + 1;
1564         int i;
1565
1566         if (index == locked_page->index && end_index == index)
1567                 return;
1568
1569         while (nr_pages > 0) {
1570                 ret = find_get_pages_contig(inode->i_mapping, index,
1571                                      min_t(unsigned long, nr_pages,
1572                                      ARRAY_SIZE(pages)), pages);
1573                 for (i = 0; i < ret; i++) {
1574                         if (pages[i] != locked_page)
1575                                 unlock_page(pages[i]);
1576                         page_cache_release(pages[i]);
1577                 }
1578                 nr_pages -= ret;
1579                 index += ret;
1580                 cond_resched();
1581         }
1582 }
1583
1584 static noinline int lock_delalloc_pages(struct inode *inode,
1585                                         struct page *locked_page,
1586                                         u64 delalloc_start,
1587                                         u64 delalloc_end)
1588 {
1589         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1590         unsigned long start_index = index;
1591         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1592         unsigned long pages_locked = 0;
1593         struct page *pages[16];
1594         unsigned long nrpages;
1595         int ret;
1596         int i;
1597
1598         /* the caller is responsible for locking the start index */
1599         if (index == locked_page->index && index == end_index)
1600                 return 0;
1601
1602         /* skip the page at the start index */
1603         nrpages = end_index - index + 1;
1604         while (nrpages > 0) {
1605                 ret = find_get_pages_contig(inode->i_mapping, index,
1606                                      min_t(unsigned long,
1607                                      nrpages, ARRAY_SIZE(pages)), pages);
1608                 if (ret == 0) {
1609                         ret = -EAGAIN;
1610                         goto done;
1611                 }
1612                 /* now we have an array of pages, lock them all */
1613                 for (i = 0; i < ret; i++) {
1614                         /*
1615                          * the caller is taking responsibility for
1616                          * locked_page
1617                          */
1618                         if (pages[i] != locked_page) {
1619                                 lock_page(pages[i]);
1620                                 if (!PageDirty(pages[i]) ||
1621                                     pages[i]->mapping != inode->i_mapping) {
1622                                         ret = -EAGAIN;
1623                                         unlock_page(pages[i]);
1624                                         page_cache_release(pages[i]);
1625                                         goto done;
1626                                 }
1627                         }
1628                         page_cache_release(pages[i]);
1629                         pages_locked++;
1630                 }
1631                 nrpages -= ret;
1632                 index += ret;
1633                 cond_resched();
1634         }
1635         ret = 0;
1636 done:
1637         if (ret && pages_locked) {
1638                 __unlock_for_delalloc(inode, locked_page,
1639                               delalloc_start,
1640                               ((u64)(start_index + pages_locked - 1)) <<
1641                               PAGE_CACHE_SHIFT);
1642         }
1643         return ret;
1644 }
1645
1646 /*
1647  * find a contiguous range of bytes in the file marked as delalloc, not
1648  * more than 'max_bytes'.  start and end are used to return the range,
1649  *
1650  * 1 is returned if we find something, 0 if nothing was in the tree
1651  */
1652 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1653                                     struct extent_io_tree *tree,
1654                                     struct page *locked_page, u64 *start,
1655                                     u64 *end, u64 max_bytes)
1656 {
1657         u64 delalloc_start;
1658         u64 delalloc_end;
1659         u64 found;
1660         struct extent_state *cached_state = NULL;
1661         int ret;
1662         int loops = 0;
1663
1664 again:
1665         /* step one, find a bunch of delalloc bytes starting at start */
1666         delalloc_start = *start;
1667         delalloc_end = 0;
1668         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1669                                     max_bytes, &cached_state);
1670         if (!found || delalloc_end <= *start) {
1671                 *start = delalloc_start;
1672                 *end = delalloc_end;
1673                 free_extent_state(cached_state);
1674                 return 0;
1675         }
1676
1677         /*
1678          * start comes from the offset of locked_page.  We have to lock
1679          * pages in order, so we can't process delalloc bytes before
1680          * locked_page
1681          */
1682         if (delalloc_start < *start)
1683                 delalloc_start = *start;
1684
1685         /*
1686          * make sure to limit the number of pages we try to lock down
1687          */
1688         if (delalloc_end + 1 - delalloc_start > max_bytes)
1689                 delalloc_end = delalloc_start + max_bytes - 1;
1690
1691         /* step two, lock all the pages after the page that has start */
1692         ret = lock_delalloc_pages(inode, locked_page,
1693                                   delalloc_start, delalloc_end);
1694         if (ret == -EAGAIN) {
1695                 /* some of the pages are gone, lets avoid looping by
1696                  * shortening the size of the delalloc range we're searching
1697                  */
1698                 free_extent_state(cached_state);
1699                 cached_state = NULL;
1700                 if (!loops) {
1701                         max_bytes = PAGE_CACHE_SIZE;
1702                         loops = 1;
1703                         goto again;
1704                 } else {
1705                         found = 0;
1706                         goto out_failed;
1707                 }
1708         }
1709         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1710
1711         /* step three, lock the state bits for the whole range */
1712         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1713
1714         /* then test to make sure it is all still delalloc */
1715         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1716                              EXTENT_DELALLOC, 1, cached_state);
1717         if (!ret) {
1718                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1719                                      &cached_state, GFP_NOFS);
1720                 __unlock_for_delalloc(inode, locked_page,
1721                               delalloc_start, delalloc_end);
1722                 cond_resched();
1723                 goto again;
1724         }
1725         free_extent_state(cached_state);
1726         *start = delalloc_start;
1727         *end = delalloc_end;
1728 out_failed:
1729         return found;
1730 }
1731
1732 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1733                                  struct page *locked_page,
1734                                  unsigned clear_bits,
1735                                  unsigned long page_ops)
1736 {
1737         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1738         int ret;
1739         struct page *pages[16];
1740         unsigned long index = start >> PAGE_CACHE_SHIFT;
1741         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1742         unsigned long nr_pages = end_index - index + 1;
1743         int i;
1744
1745         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1746         if (page_ops == 0)
1747                 return;
1748
1749         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1750                 mapping_set_error(inode->i_mapping, -EIO);
1751
1752         while (nr_pages > 0) {
1753                 ret = find_get_pages_contig(inode->i_mapping, index,
1754                                      min_t(unsigned long,
1755                                      nr_pages, ARRAY_SIZE(pages)), pages);
1756                 for (i = 0; i < ret; i++) {
1757
1758                         if (page_ops & PAGE_SET_PRIVATE2)
1759                                 SetPagePrivate2(pages[i]);
1760
1761                         if (pages[i] == locked_page) {
1762                                 page_cache_release(pages[i]);
1763                                 continue;
1764                         }
1765                         if (page_ops & PAGE_CLEAR_DIRTY)
1766                                 clear_page_dirty_for_io(pages[i]);
1767                         if (page_ops & PAGE_SET_WRITEBACK)
1768                                 set_page_writeback(pages[i]);
1769                         if (page_ops & PAGE_SET_ERROR)
1770                                 SetPageError(pages[i]);
1771                         if (page_ops & PAGE_END_WRITEBACK)
1772                                 end_page_writeback(pages[i]);
1773                         if (page_ops & PAGE_UNLOCK)
1774                                 unlock_page(pages[i]);
1775                         page_cache_release(pages[i]);
1776                 }
1777                 nr_pages -= ret;
1778                 index += ret;
1779                 cond_resched();
1780         }
1781 }
1782
1783 /*
1784  * count the number of bytes in the tree that have a given bit(s)
1785  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1786  * cached.  The total number found is returned.
1787  */
1788 u64 count_range_bits(struct extent_io_tree *tree,
1789                      u64 *start, u64 search_end, u64 max_bytes,
1790                      unsigned bits, int contig)
1791 {
1792         struct rb_node *node;
1793         struct extent_state *state;
1794         u64 cur_start = *start;
1795         u64 total_bytes = 0;
1796         u64 last = 0;
1797         int found = 0;
1798
1799         if (WARN_ON(search_end <= cur_start))
1800                 return 0;
1801
1802         spin_lock(&tree->lock);
1803         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1804                 total_bytes = tree->dirty_bytes;
1805                 goto out;
1806         }
1807         /*
1808          * this search will find all the extents that end after
1809          * our range starts.
1810          */
1811         node = tree_search(tree, cur_start);
1812         if (!node)
1813                 goto out;
1814
1815         while (1) {
1816                 state = rb_entry(node, struct extent_state, rb_node);
1817                 if (state->start > search_end)
1818                         break;
1819                 if (contig && found && state->start > last + 1)
1820                         break;
1821                 if (state->end >= cur_start && (state->state & bits) == bits) {
1822                         total_bytes += min(search_end, state->end) + 1 -
1823                                        max(cur_start, state->start);
1824                         if (total_bytes >= max_bytes)
1825                                 break;
1826                         if (!found) {
1827                                 *start = max(cur_start, state->start);
1828                                 found = 1;
1829                         }
1830                         last = state->end;
1831                 } else if (contig && found) {
1832                         break;
1833                 }
1834                 node = rb_next(node);
1835                 if (!node)
1836                         break;
1837         }
1838 out:
1839         spin_unlock(&tree->lock);
1840         return total_bytes;
1841 }
1842
1843 /*
1844  * set the private field for a given byte offset in the tree.  If there isn't
1845  * an extent_state there already, this does nothing.
1846  */
1847 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
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->private = private;
1869 out:
1870         spin_unlock(&tree->lock);
1871         return ret;
1872 }
1873
1874 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1875 {
1876         struct rb_node *node;
1877         struct extent_state *state;
1878         int ret = 0;
1879
1880         spin_lock(&tree->lock);
1881         /*
1882          * this search will find all the extents that end after
1883          * our range starts.
1884          */
1885         node = tree_search(tree, start);
1886         if (!node) {
1887                 ret = -ENOENT;
1888                 goto out;
1889         }
1890         state = rb_entry(node, struct extent_state, rb_node);
1891         if (state->start != start) {
1892                 ret = -ENOENT;
1893                 goto out;
1894         }
1895         *private = state->private;
1896 out:
1897         spin_unlock(&tree->lock);
1898         return ret;
1899 }
1900
1901 /*
1902  * searches a range in the state tree for a given mask.
1903  * If 'filled' == 1, this returns 1 only if every extent in the tree
1904  * has the bits set.  Otherwise, 1 is returned if any bit in the
1905  * range is found set.
1906  */
1907 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1908                    unsigned bits, int filled, struct extent_state *cached)
1909 {
1910         struct extent_state *state = NULL;
1911         struct rb_node *node;
1912         int bitset = 0;
1913
1914         spin_lock(&tree->lock);
1915         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1916             cached->end > start)
1917                 node = &cached->rb_node;
1918         else
1919                 node = tree_search(tree, start);
1920         while (node && start <= end) {
1921                 state = rb_entry(node, struct extent_state, rb_node);
1922
1923                 if (filled && state->start > start) {
1924                         bitset = 0;
1925                         break;
1926                 }
1927
1928                 if (state->start > end)
1929                         break;
1930
1931                 if (state->state & bits) {
1932                         bitset = 1;
1933                         if (!filled)
1934                                 break;
1935                 } else if (filled) {
1936                         bitset = 0;
1937                         break;
1938                 }
1939
1940                 if (state->end == (u64)-1)
1941                         break;
1942
1943                 start = state->end + 1;
1944                 if (start > end)
1945                         break;
1946                 node = rb_next(node);
1947                 if (!node) {
1948                         if (filled)
1949                                 bitset = 0;
1950                         break;
1951                 }
1952         }
1953         spin_unlock(&tree->lock);
1954         return bitset;
1955 }
1956
1957 /*
1958  * helper function to set a given page up to date if all the
1959  * extents in the tree for that page are up to date
1960  */
1961 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1962 {
1963         u64 start = page_offset(page);
1964         u64 end = start + PAGE_CACHE_SIZE - 1;
1965         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1966                 SetPageUptodate(page);
1967 }
1968
1969 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1970 {
1971         int ret;
1972         int err = 0;
1973         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1974
1975         set_state_private(failure_tree, rec->start, 0);
1976         ret = clear_extent_bits(failure_tree, rec->start,
1977                                 rec->start + rec->len - 1,
1978                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1979         if (ret)
1980                 err = ret;
1981
1982         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1983                                 rec->start + rec->len - 1,
1984                                 EXTENT_DAMAGED, GFP_NOFS);
1985         if (ret && !err)
1986                 err = ret;
1987
1988         kfree(rec);
1989         return err;
1990 }
1991
1992 /*
1993  * this bypasses the standard btrfs submit functions deliberately, as
1994  * the standard behavior is to write all copies in a raid setup. here we only
1995  * want to write the one bad copy. so we do the mapping for ourselves and issue
1996  * submit_bio directly.
1997  * to avoid any synchronization issues, wait for the data after writing, which
1998  * actually prevents the read that triggered the error from finishing.
1999  * currently, there can be no more than two copies of every data bit. thus,
2000  * exactly one rewrite is required.
2001  */
2002 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2003                       struct page *page, unsigned int pg_offset, int mirror_num)
2004 {
2005         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2006         struct bio *bio;
2007         struct btrfs_device *dev;
2008         u64 map_length = 0;
2009         u64 sector;
2010         struct btrfs_bio *bbio = NULL;
2011         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2012         int ret;
2013
2014         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2015         BUG_ON(!mirror_num);
2016
2017         /* we can't repair anything in raid56 yet */
2018         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2019                 return 0;
2020
2021         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2022         if (!bio)
2023                 return -EIO;
2024         bio->bi_iter.bi_size = 0;
2025         map_length = length;
2026
2027         ret = btrfs_map_block(fs_info, WRITE, logical,
2028                               &map_length, &bbio, mirror_num);
2029         if (ret) {
2030                 bio_put(bio);
2031                 return -EIO;
2032         }
2033         BUG_ON(mirror_num != bbio->mirror_num);
2034         sector = bbio->stripes[mirror_num-1].physical >> 9;
2035         bio->bi_iter.bi_sector = sector;
2036         dev = bbio->stripes[mirror_num-1].dev;
2037         btrfs_put_bbio(bbio);
2038         if (!dev || !dev->bdev || !dev->writeable) {
2039                 bio_put(bio);
2040                 return -EIO;
2041         }
2042         bio->bi_bdev = dev->bdev;
2043         bio_add_page(bio, page, length, pg_offset);
2044
2045         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio)) {
2046                 /* try to remap that extent elsewhere? */
2047                 bio_put(bio);
2048                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2049                 return -EIO;
2050         }
2051
2052         btrfs_info_rl_in_rcu(fs_info,
2053                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2054                                   btrfs_ino(inode), start,
2055                                   rcu_str_deref(dev->name), sector);
2056         bio_put(bio);
2057         return 0;
2058 }
2059
2060 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2061                          int mirror_num)
2062 {
2063         u64 start = eb->start;
2064         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2065         int ret = 0;
2066
2067         if (root->fs_info->sb->s_flags & MS_RDONLY)
2068                 return -EROFS;
2069
2070         for (i = 0; i < num_pages; i++) {
2071                 struct page *p = eb->pages[i];
2072
2073                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2074                                         PAGE_CACHE_SIZE, start, p,
2075                                         start - page_offset(p), mirror_num);
2076                 if (ret)
2077                         break;
2078                 start += PAGE_CACHE_SIZE;
2079         }
2080
2081         return ret;
2082 }
2083
2084 /*
2085  * each time an IO finishes, we do a fast check in the IO failure tree
2086  * to see if we need to process or clean up an io_failure_record
2087  */
2088 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2089                      unsigned int pg_offset)
2090 {
2091         u64 private;
2092         u64 private_failure;
2093         struct io_failure_record *failrec;
2094         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2095         struct extent_state *state;
2096         int num_copies;
2097         int ret;
2098
2099         private = 0;
2100         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2101                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2102         if (!ret)
2103                 return 0;
2104
2105         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2106                                 &private_failure);
2107         if (ret)
2108                 return 0;
2109
2110         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2111         BUG_ON(!failrec->this_mirror);
2112
2113         if (failrec->in_validation) {
2114                 /* there was no real error, just free the record */
2115                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2116                          failrec->start);
2117                 goto out;
2118         }
2119         if (fs_info->sb->s_flags & MS_RDONLY)
2120                 goto out;
2121
2122         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2123         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2124                                             failrec->start,
2125                                             EXTENT_LOCKED);
2126         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2127
2128         if (state && state->start <= failrec->start &&
2129             state->end >= failrec->start + failrec->len - 1) {
2130                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2131                                               failrec->len);
2132                 if (num_copies > 1)  {
2133                         repair_io_failure(inode, start, failrec->len,
2134                                           failrec->logical, page,
2135                                           pg_offset, failrec->failed_mirror);
2136                 }
2137         }
2138
2139 out:
2140         free_io_failure(inode, failrec);
2141
2142         return 0;
2143 }
2144
2145 /*
2146  * Can be called when
2147  * - hold extent lock
2148  * - under ordered extent
2149  * - the inode is freeing
2150  */
2151 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2152 {
2153         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2154         struct io_failure_record *failrec;
2155         struct extent_state *state, *next;
2156
2157         if (RB_EMPTY_ROOT(&failure_tree->state))
2158                 return;
2159
2160         spin_lock(&failure_tree->lock);
2161         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2162         while (state) {
2163                 if (state->start > end)
2164                         break;
2165
2166                 ASSERT(state->end <= end);
2167
2168                 next = next_state(state);
2169
2170                 failrec = (struct io_failure_record *)(unsigned long)state->private;
2171                 free_extent_state(state);
2172                 kfree(failrec);
2173
2174                 state = next;
2175         }
2176         spin_unlock(&failure_tree->lock);
2177 }
2178
2179 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2180                                 struct io_failure_record **failrec_ret)
2181 {
2182         struct io_failure_record *failrec;
2183         u64 private;
2184         struct extent_map *em;
2185         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2186         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2187         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2188         int ret;
2189         u64 logical;
2190
2191         ret = get_state_private(failure_tree, start, &private);
2192         if (ret) {
2193                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2194                 if (!failrec)
2195                         return -ENOMEM;
2196
2197                 failrec->start = start;
2198                 failrec->len = end - start + 1;
2199                 failrec->this_mirror = 0;
2200                 failrec->bio_flags = 0;
2201                 failrec->in_validation = 0;
2202
2203                 read_lock(&em_tree->lock);
2204                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2205                 if (!em) {
2206                         read_unlock(&em_tree->lock);
2207                         kfree(failrec);
2208                         return -EIO;
2209                 }
2210
2211                 if (em->start > start || em->start + em->len <= start) {
2212                         free_extent_map(em);
2213                         em = NULL;
2214                 }
2215                 read_unlock(&em_tree->lock);
2216                 if (!em) {
2217                         kfree(failrec);
2218                         return -EIO;
2219                 }
2220
2221                 logical = start - em->start;
2222                 logical = em->block_start + logical;
2223                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2224                         logical = em->block_start;
2225                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2226                         extent_set_compress_type(&failrec->bio_flags,
2227                                                  em->compress_type);
2228                 }
2229
2230                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2231                          logical, start, failrec->len);
2232
2233                 failrec->logical = logical;
2234                 free_extent_map(em);
2235
2236                 /* set the bits in the private failure tree */
2237                 ret = set_extent_bits(failure_tree, start, end,
2238                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2239                 if (ret >= 0)
2240                         ret = set_state_private(failure_tree, start,
2241                                                 (u64)(unsigned long)failrec);
2242                 /* set the bits in the inode's tree */
2243                 if (ret >= 0)
2244                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2245                                                 GFP_NOFS);
2246                 if (ret < 0) {
2247                         kfree(failrec);
2248                         return ret;
2249                 }
2250         } else {
2251                 failrec = (struct io_failure_record *)(unsigned long)private;
2252                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2253                          failrec->logical, failrec->start, failrec->len,
2254                          failrec->in_validation);
2255                 /*
2256                  * when data can be on disk more than twice, add to failrec here
2257                  * (e.g. with a list for failed_mirror) to make
2258                  * clean_io_failure() clean all those errors at once.
2259                  */
2260         }
2261
2262         *failrec_ret = failrec;
2263
2264         return 0;
2265 }
2266
2267 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2268                            struct io_failure_record *failrec, int failed_mirror)
2269 {
2270         int num_copies;
2271
2272         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2273                                       failrec->logical, failrec->len);
2274         if (num_copies == 1) {
2275                 /*
2276                  * we only have a single copy of the data, so don't bother with
2277                  * all the retry and error correction code that follows. no
2278                  * matter what the error is, it is very likely to persist.
2279                  */
2280                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2281                          num_copies, failrec->this_mirror, failed_mirror);
2282                 return 0;
2283         }
2284
2285         /*
2286          * there are two premises:
2287          *      a) deliver good data to the caller
2288          *      b) correct the bad sectors on disk
2289          */
2290         if (failed_bio->bi_vcnt > 1) {
2291                 /*
2292                  * to fulfill b), we need to know the exact failing sectors, as
2293                  * we don't want to rewrite any more than the failed ones. thus,
2294                  * we need separate read requests for the failed bio
2295                  *
2296                  * if the following BUG_ON triggers, our validation request got
2297                  * merged. we need separate requests for our algorithm to work.
2298                  */
2299                 BUG_ON(failrec->in_validation);
2300                 failrec->in_validation = 1;
2301                 failrec->this_mirror = failed_mirror;
2302         } else {
2303                 /*
2304                  * we're ready to fulfill a) and b) alongside. get a good copy
2305                  * of the failed sector and if we succeed, we have setup
2306                  * everything for repair_io_failure to do the rest for us.
2307                  */
2308                 if (failrec->in_validation) {
2309                         BUG_ON(failrec->this_mirror != failed_mirror);
2310                         failrec->in_validation = 0;
2311                         failrec->this_mirror = 0;
2312                 }
2313                 failrec->failed_mirror = failed_mirror;
2314                 failrec->this_mirror++;
2315                 if (failrec->this_mirror == failed_mirror)
2316                         failrec->this_mirror++;
2317         }
2318
2319         if (failrec->this_mirror > num_copies) {
2320                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2321                          num_copies, failrec->this_mirror, failed_mirror);
2322                 return 0;
2323         }
2324
2325         return 1;
2326 }
2327
2328
2329 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2330                                     struct io_failure_record *failrec,
2331                                     struct page *page, int pg_offset, int icsum,
2332                                     bio_end_io_t *endio_func, void *data)
2333 {
2334         struct bio *bio;
2335         struct btrfs_io_bio *btrfs_failed_bio;
2336         struct btrfs_io_bio *btrfs_bio;
2337
2338         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2339         if (!bio)
2340                 return NULL;
2341
2342         bio->bi_end_io = endio_func;
2343         bio->bi_iter.bi_sector = failrec->logical >> 9;
2344         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2345         bio->bi_iter.bi_size = 0;
2346         bio->bi_private = data;
2347
2348         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2349         if (btrfs_failed_bio->csum) {
2350                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2351                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2352
2353                 btrfs_bio = btrfs_io_bio(bio);
2354                 btrfs_bio->csum = btrfs_bio->csum_inline;
2355                 icsum *= csum_size;
2356                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2357                        csum_size);
2358         }
2359
2360         bio_add_page(bio, page, failrec->len, pg_offset);
2361
2362         return bio;
2363 }
2364
2365 /*
2366  * this is a generic handler for readpage errors (default
2367  * readpage_io_failed_hook). if other copies exist, read those and write back
2368  * good data to the failed position. does not investigate in remapping the
2369  * failed extent elsewhere, hoping the device will be smart enough to do this as
2370  * needed
2371  */
2372
2373 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2374                               struct page *page, u64 start, u64 end,
2375                               int failed_mirror)
2376 {
2377         struct io_failure_record *failrec;
2378         struct inode *inode = page->mapping->host;
2379         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2380         struct bio *bio;
2381         int read_mode;
2382         int ret;
2383
2384         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2385
2386         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2387         if (ret)
2388                 return ret;
2389
2390         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2391         if (!ret) {
2392                 free_io_failure(inode, failrec);
2393                 return -EIO;
2394         }
2395
2396         if (failed_bio->bi_vcnt > 1)
2397                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2398         else
2399                 read_mode = READ_SYNC;
2400
2401         phy_offset >>= inode->i_sb->s_blocksize_bits;
2402         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2403                                       start - page_offset(page),
2404                                       (int)phy_offset, failed_bio->bi_end_io,
2405                                       NULL);
2406         if (!bio) {
2407                 free_io_failure(inode, failrec);
2408                 return -EIO;
2409         }
2410
2411         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2412                  read_mode, failrec->this_mirror, failrec->in_validation);
2413
2414         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2415                                          failrec->this_mirror,
2416                                          failrec->bio_flags, 0);
2417         if (ret) {
2418                 free_io_failure(inode, failrec);
2419                 bio_put(bio);
2420         }
2421
2422         return ret;
2423 }
2424
2425 /* lots and lots of room for performance fixes in the end_bio funcs */
2426
2427 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2428 {
2429         int uptodate = (err == 0);
2430         struct extent_io_tree *tree;
2431         int ret = 0;
2432
2433         tree = &BTRFS_I(page->mapping->host)->io_tree;
2434
2435         if (tree->ops && tree->ops->writepage_end_io_hook) {
2436                 ret = tree->ops->writepage_end_io_hook(page, start,
2437                                                end, NULL, uptodate);
2438                 if (ret)
2439                         uptodate = 0;
2440         }
2441
2442         if (!uptodate) {
2443                 ClearPageUptodate(page);
2444                 SetPageError(page);
2445                 ret = ret < 0 ? ret : -EIO;
2446                 mapping_set_error(page->mapping, ret);
2447         }
2448 }
2449
2450 /*
2451  * after a writepage IO is done, we need to:
2452  * clear the uptodate bits on error
2453  * clear the writeback bits in the extent tree for this IO
2454  * end_page_writeback if the page has no more pending IO
2455  *
2456  * Scheduling is not allowed, so the extent state tree is expected
2457  * to have one and only one object corresponding to this IO.
2458  */
2459 static void end_bio_extent_writepage(struct bio *bio)
2460 {
2461         struct bio_vec *bvec;
2462         u64 start;
2463         u64 end;
2464         int i;
2465
2466         bio_for_each_segment_all(bvec, bio, i) {
2467                 struct page *page = bvec->bv_page;
2468
2469                 /* We always issue full-page reads, but if some block
2470                  * in a page fails to read, blk_update_request() will
2471                  * advance bv_offset and adjust bv_len to compensate.
2472                  * Print a warning for nonzero offsets, and an error
2473                  * if they don't add up to a full page.  */
2474                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2475                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2476                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2477                                    "partial page write in btrfs with offset %u and length %u",
2478                                         bvec->bv_offset, bvec->bv_len);
2479                         else
2480                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2481                                    "incomplete page write in btrfs with offset %u and "
2482                                    "length %u",
2483                                         bvec->bv_offset, bvec->bv_len);
2484                 }
2485
2486                 start = page_offset(page);
2487                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2488
2489                 end_extent_writepage(page, bio->bi_error, start, end);
2490                 end_page_writeback(page);
2491         }
2492
2493         bio_put(bio);
2494 }
2495
2496 static void
2497 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2498                               int uptodate)
2499 {
2500         struct extent_state *cached = NULL;
2501         u64 end = start + len - 1;
2502
2503         if (uptodate && tree->track_uptodate)
2504                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2505         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2506 }
2507
2508 /*
2509  * after a readpage IO is done, we need to:
2510  * clear the uptodate bits on error
2511  * set the uptodate bits if things worked
2512  * set the page up to date if all extents in the tree are uptodate
2513  * clear the lock bit in the extent tree
2514  * unlock the page if there are no other extents locked for it
2515  *
2516  * Scheduling is not allowed, so the extent state tree is expected
2517  * to have one and only one object corresponding to this IO.
2518  */
2519 static void end_bio_extent_readpage(struct bio *bio)
2520 {
2521         struct bio_vec *bvec;
2522         int uptodate = !bio->bi_error;
2523         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2524         struct extent_io_tree *tree;
2525         u64 offset = 0;
2526         u64 start;
2527         u64 end;
2528         u64 len;
2529         u64 extent_start = 0;
2530         u64 extent_len = 0;
2531         int mirror;
2532         int ret;
2533         int i;
2534
2535         bio_for_each_segment_all(bvec, bio, i) {
2536                 struct page *page = bvec->bv_page;
2537                 struct inode *inode = page->mapping->host;
2538
2539                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2540                          "mirror=%u\n", (u64)bio->bi_iter.bi_sector,
2541                          bio->bi_error, io_bio->mirror_num);
2542                 tree = &BTRFS_I(inode)->io_tree;
2543
2544                 /* We always issue full-page reads, but if some block
2545                  * in a page fails to read, blk_update_request() will
2546                  * advance bv_offset and adjust bv_len to compensate.
2547                  * Print a warning for nonzero offsets, and an error
2548                  * if they don't add up to a full page.  */
2549                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2550                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2551                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2552                                    "partial page read in btrfs with offset %u and length %u",
2553                                         bvec->bv_offset, bvec->bv_len);
2554                         else
2555                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2556                                    "incomplete page read in btrfs with offset %u and "
2557                                    "length %u",
2558                                         bvec->bv_offset, bvec->bv_len);
2559                 }
2560
2561                 start = page_offset(page);
2562                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2563                 len = bvec->bv_len;
2564
2565                 mirror = io_bio->mirror_num;
2566                 if (likely(uptodate && tree->ops &&
2567                            tree->ops->readpage_end_io_hook)) {
2568                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2569                                                               page, start, end,
2570                                                               mirror);
2571                         if (ret)
2572                                 uptodate = 0;
2573                         else
2574                                 clean_io_failure(inode, start, page, 0);
2575                 }
2576
2577                 if (likely(uptodate))
2578                         goto readpage_ok;
2579
2580                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2581                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2582                         if (!ret && !bio->bi_error)
2583                                 uptodate = 1;
2584                 } else {
2585                         /*
2586                          * The generic bio_readpage_error handles errors the
2587                          * following way: If possible, new read requests are
2588                          * created and submitted and will end up in
2589                          * end_bio_extent_readpage as well (if we're lucky, not
2590                          * in the !uptodate case). In that case it returns 0 and
2591                          * we just go on with the next page in our bio. If it
2592                          * can't handle the error it will return -EIO and we
2593                          * remain responsible for that page.
2594                          */
2595                         ret = bio_readpage_error(bio, offset, page, start, end,
2596                                                  mirror);
2597                         if (ret == 0) {
2598                                 uptodate = !bio->bi_error;
2599                                 offset += len;
2600                                 continue;
2601                         }
2602                 }
2603 readpage_ok:
2604                 if (likely(uptodate)) {
2605                         loff_t i_size = i_size_read(inode);
2606                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2607                         unsigned off;
2608
2609                         /* Zero out the end if this page straddles i_size */
2610                         off = i_size & (PAGE_CACHE_SIZE-1);
2611                         if (page->index == end_index && off)
2612                                 zero_user_segment(page, off, PAGE_CACHE_SIZE);
2613                         SetPageUptodate(page);
2614                 } else {
2615                         ClearPageUptodate(page);
2616                         SetPageError(page);
2617                 }
2618                 unlock_page(page);
2619                 offset += len;
2620
2621                 if (unlikely(!uptodate)) {
2622                         if (extent_len) {
2623                                 endio_readpage_release_extent(tree,
2624                                                               extent_start,
2625                                                               extent_len, 1);
2626                                 extent_start = 0;
2627                                 extent_len = 0;
2628                         }
2629                         endio_readpage_release_extent(tree, start,
2630                                                       end - start + 1, 0);
2631                 } else if (!extent_len) {
2632                         extent_start = start;
2633                         extent_len = end + 1 - start;
2634                 } else if (extent_start + extent_len == start) {
2635                         extent_len += end + 1 - start;
2636                 } else {
2637                         endio_readpage_release_extent(tree, extent_start,
2638                                                       extent_len, uptodate);
2639                         extent_start = start;
2640                         extent_len = end + 1 - start;
2641                 }
2642         }
2643
2644         if (extent_len)
2645                 endio_readpage_release_extent(tree, extent_start, extent_len,
2646                                               uptodate);
2647         if (io_bio->end_io)
2648                 io_bio->end_io(io_bio, bio->bi_error);
2649         bio_put(bio);
2650 }
2651
2652 /*
2653  * this allocates from the btrfs_bioset.  We're returning a bio right now
2654  * but you can call btrfs_io_bio for the appropriate container_of magic
2655  */
2656 struct bio *
2657 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2658                 gfp_t gfp_flags)
2659 {
2660         struct btrfs_io_bio *btrfs_bio;
2661         struct bio *bio;
2662
2663         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2664
2665         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2666                 while (!bio && (nr_vecs /= 2)) {
2667                         bio = bio_alloc_bioset(gfp_flags,
2668                                                nr_vecs, btrfs_bioset);
2669                 }
2670         }
2671
2672         if (bio) {
2673                 bio->bi_bdev = bdev;
2674                 bio->bi_iter.bi_sector = first_sector;
2675                 btrfs_bio = btrfs_io_bio(bio);
2676                 btrfs_bio->csum = NULL;
2677                 btrfs_bio->csum_allocated = NULL;
2678                 btrfs_bio->end_io = NULL;
2679         }
2680         return bio;
2681 }
2682
2683 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2684 {
2685         struct btrfs_io_bio *btrfs_bio;
2686         struct bio *new;
2687
2688         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2689         if (new) {
2690                 btrfs_bio = btrfs_io_bio(new);
2691                 btrfs_bio->csum = NULL;
2692                 btrfs_bio->csum_allocated = NULL;
2693                 btrfs_bio->end_io = NULL;
2694
2695 #ifdef CONFIG_BLK_CGROUP
2696                 /* FIXME, put this into bio_clone_bioset */
2697                 if (bio->bi_css)
2698                         bio_associate_blkcg(new, bio->bi_css);
2699 #endif
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(int rw, struct bio *bio,
2722                                        int mirror_num, 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
2734         bio_get(bio);
2735
2736         if (tree->ops && tree->ops->submit_bio_hook)
2737                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2738                                            mirror_num, bio_flags, start);
2739         else
2740                 btrfsic_submit_bio(rw, bio);
2741
2742         bio_put(bio);
2743         return ret;
2744 }
2745
2746 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2747                      unsigned long offset, size_t size, struct bio *bio,
2748                      unsigned long bio_flags)
2749 {
2750         int ret = 0;
2751         if (tree->ops && tree->ops->merge_bio_hook)
2752                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2753                                                 bio_flags);
2754         BUG_ON(ret < 0);
2755         return ret;
2756
2757 }
2758
2759 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2760                               struct writeback_control *wbc,
2761                               struct page *page, sector_t sector,
2762                               size_t size, unsigned long offset,
2763                               struct block_device *bdev,
2764                               struct bio **bio_ret,
2765                               unsigned long max_pages,
2766                               bio_end_io_t end_io_func,
2767                               int mirror_num,
2768                               unsigned long prev_bio_flags,
2769                               unsigned long bio_flags,
2770                               bool force_bio_submit)
2771 {
2772         int ret = 0;
2773         struct bio *bio;
2774         int contig = 0;
2775         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2776         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2777
2778         if (bio_ret && *bio_ret) {
2779                 bio = *bio_ret;
2780                 if (old_compressed)
2781                         contig = bio->bi_iter.bi_sector == sector;
2782                 else
2783                         contig = bio_end_sector(bio) == sector;
2784
2785                 if (prev_bio_flags != bio_flags || !contig ||
2786                     force_bio_submit ||
2787                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2788                     bio_add_page(bio, page, page_size, offset) < page_size) {
2789                         ret = submit_one_bio(rw, bio, mirror_num,
2790                                              prev_bio_flags);
2791                         if (ret < 0) {
2792                                 *bio_ret = NULL;
2793                                 return ret;
2794                         }
2795                         bio = NULL;
2796                 } else {
2797                         if (wbc)
2798                                 wbc_account_io(wbc, page, page_size);
2799                         return 0;
2800                 }
2801         }
2802
2803         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2804                         GFP_NOFS | __GFP_HIGH);
2805         if (!bio)
2806                 return -ENOMEM;
2807
2808         bio_add_page(bio, page, page_size, offset);
2809         bio->bi_end_io = end_io_func;
2810         bio->bi_private = tree;
2811         if (wbc) {
2812                 wbc_init_bio(wbc, bio);
2813                 wbc_account_io(wbc, page, page_size);
2814         }
2815
2816         if (bio_ret)
2817                 *bio_ret = bio;
2818         else
2819                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2820
2821         return ret;
2822 }
2823
2824 static void attach_extent_buffer_page(struct extent_buffer *eb,
2825                                       struct page *page)
2826 {
2827         if (!PagePrivate(page)) {
2828                 SetPagePrivate(page);
2829                 page_cache_get(page);
2830                 set_page_private(page, (unsigned long)eb);
2831         } else {
2832                 WARN_ON(page->private != (unsigned long)eb);
2833         }
2834 }
2835
2836 void set_page_extent_mapped(struct page *page)
2837 {
2838         if (!PagePrivate(page)) {
2839                 SetPagePrivate(page);
2840                 page_cache_get(page);
2841                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2842         }
2843 }
2844
2845 static struct extent_map *
2846 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2847                  u64 start, u64 len, get_extent_t *get_extent,
2848                  struct extent_map **em_cached)
2849 {
2850         struct extent_map *em;
2851
2852         if (em_cached && *em_cached) {
2853                 em = *em_cached;
2854                 if (extent_map_in_tree(em) && start >= em->start &&
2855                     start < extent_map_end(em)) {
2856                         atomic_inc(&em->refs);
2857                         return em;
2858                 }
2859
2860                 free_extent_map(em);
2861                 *em_cached = NULL;
2862         }
2863
2864         em = get_extent(inode, page, pg_offset, start, len, 0);
2865         if (em_cached && !IS_ERR_OR_NULL(em)) {
2866                 BUG_ON(*em_cached);
2867                 atomic_inc(&em->refs);
2868                 *em_cached = em;
2869         }
2870         return em;
2871 }
2872 /*
2873  * basic readpage implementation.  Locked extent state structs are inserted
2874  * into the tree that are removed when the IO is done (by the end_io
2875  * handlers)
2876  * XXX JDM: This needs looking at to ensure proper page locking
2877  */
2878 static int __do_readpage(struct extent_io_tree *tree,
2879                          struct page *page,
2880                          get_extent_t *get_extent,
2881                          struct extent_map **em_cached,
2882                          struct bio **bio, int mirror_num,
2883                          unsigned long *bio_flags, int rw,
2884                          u64 *prev_em_start)
2885 {
2886         struct inode *inode = page->mapping->host;
2887         u64 start = page_offset(page);
2888         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2889         u64 end;
2890         u64 cur = start;
2891         u64 extent_offset;
2892         u64 last_byte = i_size_read(inode);
2893         u64 block_start;
2894         u64 cur_end;
2895         sector_t sector;
2896         struct extent_map *em;
2897         struct block_device *bdev;
2898         int ret;
2899         int nr = 0;
2900         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2901         size_t pg_offset = 0;
2902         size_t iosize;
2903         size_t disk_io_size;
2904         size_t blocksize = inode->i_sb->s_blocksize;
2905         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2906
2907         set_page_extent_mapped(page);
2908
2909         end = page_end;
2910         if (!PageUptodate(page)) {
2911                 if (cleancache_get_page(page) == 0) {
2912                         BUG_ON(blocksize != PAGE_SIZE);
2913                         unlock_extent(tree, start, end);
2914                         goto out;
2915                 }
2916         }
2917
2918         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2919                 char *userpage;
2920                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2921
2922                 if (zero_offset) {
2923                         iosize = PAGE_CACHE_SIZE - zero_offset;
2924                         userpage = kmap_atomic(page);
2925                         memset(userpage + zero_offset, 0, iosize);
2926                         flush_dcache_page(page);
2927                         kunmap_atomic(userpage);
2928                 }
2929         }
2930         while (cur <= end) {
2931                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2932                 bool force_bio_submit = false;
2933
2934                 if (cur >= last_byte) {
2935                         char *userpage;
2936                         struct extent_state *cached = NULL;
2937
2938                         iosize = PAGE_CACHE_SIZE - pg_offset;
2939                         userpage = kmap_atomic(page);
2940                         memset(userpage + pg_offset, 0, iosize);
2941                         flush_dcache_page(page);
2942                         kunmap_atomic(userpage);
2943                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2944                                             &cached, GFP_NOFS);
2945                         if (!parent_locked)
2946                                 unlock_extent_cached(tree, cur,
2947                                                      cur + iosize - 1,
2948                                                      &cached, GFP_NOFS);
2949                         break;
2950                 }
2951                 em = __get_extent_map(inode, page, pg_offset, cur,
2952                                       end - cur + 1, get_extent, em_cached);
2953                 if (IS_ERR_OR_NULL(em)) {
2954                         SetPageError(page);
2955                         if (!parent_locked)
2956                                 unlock_extent(tree, cur, end);
2957                         break;
2958                 }
2959                 extent_offset = cur - em->start;
2960                 BUG_ON(extent_map_end(em) <= cur);
2961                 BUG_ON(end < cur);
2962
2963                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2964                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2965                         extent_set_compress_type(&this_bio_flag,
2966                                                  em->compress_type);
2967                 }
2968
2969                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2970                 cur_end = min(extent_map_end(em) - 1, end);
2971                 iosize = ALIGN(iosize, blocksize);
2972                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2973                         disk_io_size = em->block_len;
2974                         sector = em->block_start >> 9;
2975                 } else {
2976                         sector = (em->block_start + extent_offset) >> 9;
2977                         disk_io_size = iosize;
2978                 }
2979                 bdev = em->bdev;
2980                 block_start = em->block_start;
2981                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2982                         block_start = EXTENT_MAP_HOLE;
2983
2984                 /*
2985                  * If we have a file range that points to a compressed extent
2986                  * and it's followed by a consecutive file range that points to
2987                  * to the same compressed extent (possibly with a different
2988                  * offset and/or length, so it either points to the whole extent
2989                  * or only part of it), we must make sure we do not submit a
2990                  * single bio to populate the pages for the 2 ranges because
2991                  * this makes the compressed extent read zero out the pages
2992                  * belonging to the 2nd range. Imagine the following scenario:
2993                  *
2994                  *  File layout
2995                  *  [0 - 8K]                     [8K - 24K]
2996                  *    |                               |
2997                  *    |                               |
2998                  * points to extent X,         points to extent X,
2999                  * offset 4K, length of 8K     offset 0, length 16K
3000                  *
3001                  * [extent X, compressed length = 4K uncompressed length = 16K]
3002                  *
3003                  * If the bio to read the compressed extent covers both ranges,
3004                  * it will decompress extent X into the pages belonging to the
3005                  * first range and then it will stop, zeroing out the remaining
3006                  * pages that belong to the other range that points to extent X.
3007                  * So here we make sure we submit 2 bios, one for the first
3008                  * range and another one for the third range. Both will target
3009                  * the same physical extent from disk, but we can't currently
3010                  * make the compressed bio endio callback populate the pages
3011                  * for both ranges because each compressed bio is tightly
3012                  * coupled with a single extent map, and each range can have
3013                  * an extent map with a different offset value relative to the
3014                  * uncompressed data of our extent and different lengths. This
3015                  * is a corner case so we prioritize correctness over
3016                  * non-optimal behavior (submitting 2 bios for the same extent).
3017                  */
3018                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3019                     prev_em_start && *prev_em_start != (u64)-1 &&
3020                     *prev_em_start != em->orig_start)
3021                         force_bio_submit = true;
3022
3023                 if (prev_em_start)
3024                         *prev_em_start = em->orig_start;
3025
3026                 free_extent_map(em);
3027                 em = NULL;
3028
3029                 /* we've found a hole, just zero and go on */
3030                 if (block_start == EXTENT_MAP_HOLE) {
3031                         char *userpage;
3032                         struct extent_state *cached = NULL;
3033
3034                         userpage = kmap_atomic(page);
3035                         memset(userpage + pg_offset, 0, iosize);
3036                         flush_dcache_page(page);
3037                         kunmap_atomic(userpage);
3038
3039                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3040                                             &cached, GFP_NOFS);
3041                         if (parent_locked)
3042                                 free_extent_state(cached);
3043                         else
3044                                 unlock_extent_cached(tree, cur,
3045                                                      cur + iosize - 1,
3046                                                      &cached, GFP_NOFS);
3047                         cur = cur + iosize;
3048                         pg_offset += iosize;
3049                         continue;
3050                 }
3051                 /* the get_extent function already copied into the page */
3052                 if (test_range_bit(tree, cur, cur_end,
3053                                    EXTENT_UPTODATE, 1, NULL)) {
3054                         check_page_uptodate(tree, page);
3055                         if (!parent_locked)
3056                                 unlock_extent(tree, cur, cur + iosize - 1);
3057                         cur = cur + iosize;
3058                         pg_offset += iosize;
3059                         continue;
3060                 }
3061                 /* we have an inline extent but it didn't get marked up
3062                  * to date.  Error out
3063                  */
3064                 if (block_start == EXTENT_MAP_INLINE) {
3065                         SetPageError(page);
3066                         if (!parent_locked)
3067                                 unlock_extent(tree, cur, cur + iosize - 1);
3068                         cur = cur + iosize;
3069                         pg_offset += iosize;
3070                         continue;
3071                 }
3072
3073                 pnr -= page->index;
3074                 ret = submit_extent_page(rw, tree, NULL, page,
3075                                          sector, disk_io_size, pg_offset,
3076                                          bdev, bio, pnr,
3077                                          end_bio_extent_readpage, mirror_num,
3078                                          *bio_flags,
3079                                          this_bio_flag,
3080                                          force_bio_submit);
3081                 if (!ret) {
3082                         nr++;
3083                         *bio_flags = this_bio_flag;
3084                 } else {
3085                         SetPageError(page);
3086                         if (!parent_locked)
3087                                 unlock_extent(tree, cur, cur + iosize - 1);
3088                 }
3089                 cur = cur + iosize;
3090                 pg_offset += iosize;
3091         }
3092 out:
3093         if (!nr) {
3094                 if (!PageError(page))
3095                         SetPageUptodate(page);
3096                 unlock_page(page);
3097         }
3098         return 0;
3099 }
3100
3101 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3102                                              struct page *pages[], int nr_pages,
3103                                              u64 start, u64 end,
3104                                              get_extent_t *get_extent,
3105                                              struct extent_map **em_cached,
3106                                              struct bio **bio, int mirror_num,
3107                                              unsigned long *bio_flags, int rw,
3108                                              u64 *prev_em_start)
3109 {
3110         struct inode *inode;
3111         struct btrfs_ordered_extent *ordered;
3112         int index;
3113
3114         inode = pages[0]->mapping->host;
3115         while (1) {
3116                 lock_extent(tree, start, end);
3117                 ordered = btrfs_lookup_ordered_range(inode, start,
3118                                                      end - start + 1);
3119                 if (!ordered)
3120                         break;
3121                 unlock_extent(tree, start, end);
3122                 btrfs_start_ordered_extent(inode, ordered, 1);
3123                 btrfs_put_ordered_extent(ordered);
3124         }
3125
3126         for (index = 0; index < nr_pages; index++) {
3127                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3128                               mirror_num, bio_flags, rw, prev_em_start);
3129                 page_cache_release(pages[index]);
3130         }
3131 }
3132
3133 static void __extent_readpages(struct extent_io_tree *tree,
3134                                struct page *pages[],
3135                                int nr_pages, get_extent_t *get_extent,
3136                                struct extent_map **em_cached,
3137                                struct bio **bio, int mirror_num,
3138                                unsigned long *bio_flags, int rw,
3139                                u64 *prev_em_start)
3140 {
3141         u64 start = 0;
3142         u64 end = 0;
3143         u64 page_start;
3144         int index;
3145         int first_index = 0;
3146
3147         for (index = 0; index < nr_pages; index++) {
3148                 page_start = page_offset(pages[index]);
3149                 if (!end) {
3150                         start = page_start;
3151                         end = start + PAGE_CACHE_SIZE - 1;
3152                         first_index = index;
3153                 } else if (end + 1 == page_start) {
3154                         end += PAGE_CACHE_SIZE;
3155                 } else {
3156                         __do_contiguous_readpages(tree, &pages[first_index],
3157                                                   index - first_index, start,
3158                                                   end, get_extent, em_cached,
3159                                                   bio, mirror_num, bio_flags,
3160                                                   rw, prev_em_start);
3161                         start = page_start;
3162                         end = start + PAGE_CACHE_SIZE - 1;
3163                         first_index = index;
3164                 }
3165         }
3166
3167         if (end)
3168                 __do_contiguous_readpages(tree, &pages[first_index],
3169                                           index - first_index, start,
3170                                           end, get_extent, em_cached, bio,
3171                                           mirror_num, bio_flags, rw,
3172                                           prev_em_start);
3173 }
3174
3175 static int __extent_read_full_page(struct extent_io_tree *tree,
3176                                    struct page *page,
3177                                    get_extent_t *get_extent,
3178                                    struct bio **bio, int mirror_num,
3179                                    unsigned long *bio_flags, int rw)
3180 {
3181         struct inode *inode = page->mapping->host;
3182         struct btrfs_ordered_extent *ordered;
3183         u64 start = page_offset(page);
3184         u64 end = start + PAGE_CACHE_SIZE - 1;
3185         int ret;
3186
3187         while (1) {
3188                 lock_extent(tree, start, end);
3189                 ordered = btrfs_lookup_ordered_extent(inode, start);
3190                 if (!ordered)
3191                         break;
3192                 unlock_extent(tree, start, end);
3193                 btrfs_start_ordered_extent(inode, ordered, 1);
3194                 btrfs_put_ordered_extent(ordered);
3195         }
3196
3197         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3198                             bio_flags, rw, NULL);
3199         return ret;
3200 }
3201
3202 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3203                             get_extent_t *get_extent, int mirror_num)
3204 {
3205         struct bio *bio = NULL;
3206         unsigned long bio_flags = 0;
3207         int ret;
3208
3209         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3210                                       &bio_flags, READ);
3211         if (bio)
3212                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3213         return ret;
3214 }
3215
3216 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3217                                  get_extent_t *get_extent, int mirror_num)
3218 {
3219         struct bio *bio = NULL;
3220         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3221         int ret;
3222
3223         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3224                             &bio_flags, READ, NULL);
3225         if (bio)
3226                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3227         return ret;
3228 }
3229
3230 static noinline void update_nr_written(struct page *page,
3231                                       struct writeback_control *wbc,
3232                                       unsigned long nr_written)
3233 {
3234         wbc->nr_to_write -= nr_written;
3235         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3236             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3237                 page->mapping->writeback_index = page->index + nr_written;
3238 }
3239
3240 /*
3241  * helper for __extent_writepage, doing all of the delayed allocation setup.
3242  *
3243  * This returns 1 if our fill_delalloc function did all the work required
3244  * to write the page (copy into inline extent).  In this case the IO has
3245  * been started and the page is already unlocked.
3246  *
3247  * This returns 0 if all went well (page still locked)
3248  * This returns < 0 if there were errors (page still locked)
3249  */
3250 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3251                               struct page *page, struct writeback_control *wbc,
3252                               struct extent_page_data *epd,
3253                               u64 delalloc_start,
3254                               unsigned long *nr_written)
3255 {
3256         struct extent_io_tree *tree = epd->tree;
3257         u64 page_end = delalloc_start + PAGE_CACHE_SIZE - 1;
3258         u64 nr_delalloc;
3259         u64 delalloc_to_write = 0;
3260         u64 delalloc_end = 0;
3261         int ret;
3262         int page_started = 0;
3263
3264         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3265                 return 0;
3266
3267         while (delalloc_end < page_end) {
3268                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3269                                                page,
3270                                                &delalloc_start,
3271                                                &delalloc_end,
3272                                                BTRFS_MAX_EXTENT_SIZE);
3273                 if (nr_delalloc == 0) {
3274                         delalloc_start = delalloc_end + 1;
3275                         continue;
3276                 }
3277                 ret = tree->ops->fill_delalloc(inode, page,
3278                                                delalloc_start,
3279                                                delalloc_end,
3280                                                &page_started,
3281                                                nr_written);
3282                 /* File system has been set read-only */
3283                 if (ret) {
3284                         SetPageError(page);
3285                         /* fill_delalloc should be return < 0 for error
3286                          * but just in case, we use > 0 here meaning the
3287                          * IO is started, so we don't want to return > 0
3288                          * unless things are going well.
3289                          */
3290                         ret = ret < 0 ? ret : -EIO;
3291                         goto done;
3292                 }
3293                 /*
3294                  * delalloc_end is already one less than the total
3295                  * length, so we don't subtract one from
3296                  * PAGE_CACHE_SIZE
3297                  */
3298                 delalloc_to_write += (delalloc_end - delalloc_start +
3299                                       PAGE_CACHE_SIZE) >>
3300                                       PAGE_CACHE_SHIFT;
3301                 delalloc_start = delalloc_end + 1;
3302         }
3303         if (wbc->nr_to_write < delalloc_to_write) {
3304                 int thresh = 8192;
3305
3306                 if (delalloc_to_write < thresh * 2)
3307                         thresh = delalloc_to_write;
3308                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3309                                          thresh);
3310         }
3311
3312         /* did the fill delalloc function already unlock and start
3313          * the IO?
3314          */
3315         if (page_started) {
3316                 /*
3317                  * we've unlocked the page, so we can't update
3318                  * the mapping's writeback index, just update
3319                  * nr_to_write.
3320                  */
3321                 wbc->nr_to_write -= *nr_written;
3322                 return 1;
3323         }
3324
3325         ret = 0;
3326
3327 done:
3328         return ret;
3329 }
3330
3331 /*
3332  * helper for __extent_writepage.  This calls the writepage start hooks,
3333  * and does the loop to map the page into extents and bios.
3334  *
3335  * We return 1 if the IO is started and the page is unlocked,
3336  * 0 if all went well (page still locked)
3337  * < 0 if there were errors (page still locked)
3338  */
3339 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3340                                  struct page *page,
3341                                  struct writeback_control *wbc,
3342                                  struct extent_page_data *epd,
3343                                  loff_t i_size,
3344                                  unsigned long nr_written,
3345                                  int write_flags, int *nr_ret)
3346 {
3347         struct extent_io_tree *tree = epd->tree;
3348         u64 start = page_offset(page);
3349         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3350         u64 end;
3351         u64 cur = start;
3352         u64 extent_offset;
3353         u64 block_start;
3354         u64 iosize;
3355         sector_t sector;
3356         struct extent_state *cached_state = NULL;
3357         struct extent_map *em;
3358         struct block_device *bdev;
3359         size_t pg_offset = 0;
3360         size_t blocksize;
3361         int ret = 0;
3362         int nr = 0;
3363         bool compressed;
3364
3365         if (tree->ops && tree->ops->writepage_start_hook) {
3366                 ret = tree->ops->writepage_start_hook(page, start,
3367                                                       page_end);
3368                 if (ret) {
3369                         /* Fixup worker will requeue */
3370                         if (ret == -EBUSY)
3371                                 wbc->pages_skipped++;
3372                         else
3373                                 redirty_page_for_writepage(wbc, page);
3374
3375                         update_nr_written(page, wbc, nr_written);
3376                         unlock_page(page);
3377                         ret = 1;
3378                         goto done_unlocked;
3379                 }
3380         }
3381
3382         /*
3383          * we don't want to touch the inode after unlocking the page,
3384          * so we update the mapping writeback index now
3385          */
3386         update_nr_written(page, wbc, nr_written + 1);
3387
3388         end = page_end;
3389         if (i_size <= start) {
3390                 if (tree->ops && tree->ops->writepage_end_io_hook)
3391                         tree->ops->writepage_end_io_hook(page, start,
3392                                                          page_end, NULL, 1);
3393                 goto done;
3394         }
3395
3396         blocksize = inode->i_sb->s_blocksize;
3397
3398         while (cur <= end) {
3399                 u64 em_end;
3400                 if (cur >= i_size) {
3401                         if (tree->ops && tree->ops->writepage_end_io_hook)
3402                                 tree->ops->writepage_end_io_hook(page, cur,
3403                                                          page_end, NULL, 1);
3404                         break;
3405                 }
3406                 em = epd->get_extent(inode, page, pg_offset, cur,
3407                                      end - cur + 1, 1);
3408                 if (IS_ERR_OR_NULL(em)) {
3409                         SetPageError(page);
3410                         ret = PTR_ERR_OR_ZERO(em);
3411                         break;
3412                 }
3413
3414                 extent_offset = cur - em->start;
3415                 em_end = extent_map_end(em);
3416                 BUG_ON(em_end <= cur);
3417                 BUG_ON(end < cur);
3418                 iosize = min(em_end - cur, end - cur + 1);
3419                 iosize = ALIGN(iosize, blocksize);
3420                 sector = (em->block_start + extent_offset) >> 9;
3421                 bdev = em->bdev;
3422                 block_start = em->block_start;
3423                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3424                 free_extent_map(em);
3425                 em = NULL;
3426
3427                 /*
3428                  * compressed and inline extents are written through other
3429                  * paths in the FS
3430                  */
3431                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3432                     block_start == EXTENT_MAP_INLINE) {
3433                         /*
3434                          * end_io notification does not happen here for
3435                          * compressed extents
3436                          */
3437                         if (!compressed && tree->ops &&
3438                             tree->ops->writepage_end_io_hook)
3439                                 tree->ops->writepage_end_io_hook(page, cur,
3440                                                          cur + iosize - 1,
3441                                                          NULL, 1);
3442                         else if (compressed) {
3443                                 /* we don't want to end_page_writeback on
3444                                  * a compressed extent.  this happens
3445                                  * elsewhere
3446                                  */
3447                                 nr++;
3448                         }
3449
3450                         cur += iosize;
3451                         pg_offset += iosize;
3452                         continue;
3453                 }
3454
3455                 if (tree->ops && tree->ops->writepage_io_hook) {
3456                         ret = tree->ops->writepage_io_hook(page, cur,
3457                                                 cur + iosize - 1);
3458                 } else {
3459                         ret = 0;
3460                 }
3461                 if (ret) {
3462                         SetPageError(page);
3463                 } else {
3464                         unsigned long max_nr = (i_size >> PAGE_CACHE_SHIFT) + 1;
3465
3466                         set_range_writeback(tree, cur, cur + iosize - 1);
3467                         if (!PageWriteback(page)) {
3468                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
3469                                            "page %lu not writeback, cur %llu end %llu",
3470                                        page->index, cur, end);
3471                         }
3472
3473                         ret = submit_extent_page(write_flags, tree, wbc, page,
3474                                                  sector, iosize, pg_offset,
3475                                                  bdev, &epd->bio, max_nr,
3476                                                  end_bio_extent_writepage,
3477                                                  0, 0, 0, false);
3478                         if (ret)
3479                                 SetPageError(page);
3480                 }
3481                 cur = cur + iosize;
3482                 pg_offset += iosize;
3483                 nr++;
3484         }
3485 done:
3486         *nr_ret = nr;
3487
3488 done_unlocked:
3489
3490         /* drop our reference on any cached states */
3491         free_extent_state(cached_state);
3492         return ret;
3493 }
3494
3495 /*
3496  * the writepage semantics are similar to regular writepage.  extent
3497  * records are inserted to lock ranges in the tree, and as dirty areas
3498  * are found, they are marked writeback.  Then the lock bits are removed
3499  * and the end_io handler clears the writeback ranges
3500  */
3501 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3502                               void *data)
3503 {
3504         struct inode *inode = page->mapping->host;
3505         struct extent_page_data *epd = data;
3506         u64 start = page_offset(page);
3507         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3508         int ret;
3509         int nr = 0;
3510         size_t pg_offset = 0;
3511         loff_t i_size = i_size_read(inode);
3512         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3513         int write_flags;
3514         unsigned long nr_written = 0;
3515
3516         if (wbc->sync_mode == WB_SYNC_ALL)
3517                 write_flags = WRITE_SYNC;
3518         else
3519                 write_flags = WRITE;
3520
3521         trace___extent_writepage(page, inode, wbc);
3522
3523         WARN_ON(!PageLocked(page));
3524
3525         ClearPageError(page);
3526
3527         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3528         if (page->index > end_index ||
3529            (page->index == end_index && !pg_offset)) {
3530                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3531                 unlock_page(page);
3532                 return 0;
3533         }
3534
3535         if (page->index == end_index) {
3536                 char *userpage;
3537
3538                 userpage = kmap_atomic(page);
3539                 memset(userpage + pg_offset, 0,
3540                        PAGE_CACHE_SIZE - pg_offset);
3541                 kunmap_atomic(userpage);
3542                 flush_dcache_page(page);
3543         }
3544
3545         pg_offset = 0;
3546
3547         set_page_extent_mapped(page);
3548
3549         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3550         if (ret == 1)
3551                 goto done_unlocked;
3552         if (ret)
3553                 goto done;
3554
3555         ret = __extent_writepage_io(inode, page, wbc, epd,
3556                                     i_size, nr_written, write_flags, &nr);
3557         if (ret == 1)
3558                 goto done_unlocked;
3559
3560 done:
3561         if (nr == 0) {
3562                 /* make sure the mapping tag for page dirty gets cleared */
3563                 set_page_writeback(page);
3564                 end_page_writeback(page);
3565         }
3566         if (PageError(page)) {
3567                 ret = ret < 0 ? ret : -EIO;
3568                 end_extent_writepage(page, ret, start, page_end);
3569         }
3570         unlock_page(page);
3571         return ret;
3572
3573 done_unlocked:
3574         return 0;
3575 }
3576
3577 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3578 {
3579         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3580                        TASK_UNINTERRUPTIBLE);
3581 }
3582
3583 static noinline_for_stack int
3584 lock_extent_buffer_for_io(struct extent_buffer *eb,
3585                           struct btrfs_fs_info *fs_info,
3586                           struct extent_page_data *epd)
3587 {
3588         unsigned long i, num_pages;
3589         int flush = 0;
3590         int ret = 0;
3591
3592         if (!btrfs_try_tree_write_lock(eb)) {
3593                 flush = 1;
3594                 flush_write_bio(epd);
3595                 btrfs_tree_lock(eb);
3596         }
3597
3598         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3599                 btrfs_tree_unlock(eb);
3600                 if (!epd->sync_io)
3601                         return 0;
3602                 if (!flush) {
3603                         flush_write_bio(epd);
3604                         flush = 1;
3605                 }
3606                 while (1) {
3607                         wait_on_extent_buffer_writeback(eb);
3608                         btrfs_tree_lock(eb);
3609                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3610                                 break;
3611                         btrfs_tree_unlock(eb);
3612                 }
3613         }
3614
3615         /*
3616          * We need to do this to prevent races in people who check if the eb is
3617          * under IO since we can end up having no IO bits set for a short period
3618          * of time.
3619          */
3620         spin_lock(&eb->refs_lock);
3621         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3622                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3623                 spin_unlock(&eb->refs_lock);
3624                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3625                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3626                                      -eb->len,
3627                                      fs_info->dirty_metadata_batch);
3628                 ret = 1;
3629         } else {
3630                 spin_unlock(&eb->refs_lock);
3631         }
3632
3633         btrfs_tree_unlock(eb);
3634
3635         if (!ret)
3636                 return ret;
3637
3638         num_pages = num_extent_pages(eb->start, eb->len);
3639         for (i = 0; i < num_pages; i++) {
3640                 struct page *p = eb->pages[i];
3641
3642                 if (!trylock_page(p)) {
3643                         if (!flush) {
3644                                 flush_write_bio(epd);
3645                                 flush = 1;
3646                         }
3647                         lock_page(p);
3648                 }
3649         }
3650
3651         return ret;
3652 }
3653
3654 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3655 {
3656         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3657         smp_mb__after_atomic();
3658         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3659 }
3660
3661 static void set_btree_ioerr(struct page *page)
3662 {
3663         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3664         struct btrfs_inode *btree_ino = BTRFS_I(eb->fs_info->btree_inode);
3665
3666         SetPageError(page);
3667         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3668                 return;
3669
3670         /*
3671          * If writeback for a btree extent that doesn't belong to a log tree
3672          * failed, increment the counter transaction->eb_write_errors.
3673          * We do this because while the transaction is running and before it's
3674          * committing (when we call filemap_fdata[write|wait]_range against
3675          * the btree inode), we might have
3676          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3677          * returns an error or an error happens during writeback, when we're
3678          * committing the transaction we wouldn't know about it, since the pages
3679          * can be no longer dirty nor marked anymore for writeback (if a
3680          * subsequent modification to the extent buffer didn't happen before the
3681          * transaction commit), which makes filemap_fdata[write|wait]_range not
3682          * able to find the pages tagged with SetPageError at transaction
3683          * commit time. So if this happens we must abort the transaction,
3684          * otherwise we commit a super block with btree roots that point to
3685          * btree nodes/leafs whose content on disk is invalid - either garbage
3686          * or the content of some node/leaf from a past generation that got
3687          * cowed or deleted and is no longer valid.
3688          *
3689          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3690          * not be enough - we need to distinguish between log tree extents vs
3691          * non-log tree extents, and the next filemap_fdatawait_range() call
3692          * will catch and clear such errors in the mapping - and that call might
3693          * be from a log sync and not from a transaction commit. Also, checking
3694          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3695          * not done and would not be reliable - the eb might have been released
3696          * from memory and reading it back again means that flag would not be
3697          * set (since it's a runtime flag, not persisted on disk).
3698          *
3699          * Using the flags below in the btree inode also makes us achieve the
3700          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3701          * writeback for all dirty pages and before filemap_fdatawait_range()
3702          * is called, the writeback for all dirty pages had already finished
3703          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3704          * filemap_fdatawait_range() would return success, as it could not know
3705          * that writeback errors happened (the pages were no longer tagged for
3706          * writeback).
3707          */
3708         switch (eb->log_index) {
3709         case -1:
3710                 set_bit(BTRFS_INODE_BTREE_ERR, &btree_ino->runtime_flags);
3711                 break;
3712         case 0:
3713                 set_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
3714                 break;
3715         case 1:
3716                 set_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
3717                 break;
3718         default:
3719                 BUG(); /* unexpected, logic error */
3720         }
3721 }
3722
3723 static void end_bio_extent_buffer_writepage(struct bio *bio)
3724 {
3725         struct bio_vec *bvec;
3726         struct extent_buffer *eb;
3727         int i, done;
3728
3729         bio_for_each_segment_all(bvec, bio, i) {
3730                 struct page *page = bvec->bv_page;
3731
3732                 eb = (struct extent_buffer *)page->private;
3733                 BUG_ON(!eb);
3734                 done = atomic_dec_and_test(&eb->io_pages);
3735
3736                 if (bio->bi_error ||
3737                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3738                         ClearPageUptodate(page);
3739                         set_btree_ioerr(page);
3740                 }
3741
3742                 end_page_writeback(page);
3743
3744                 if (!done)
3745                         continue;
3746
3747                 end_extent_buffer_writeback(eb);
3748         }
3749
3750         bio_put(bio);
3751 }
3752
3753 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3754                         struct btrfs_fs_info *fs_info,
3755                         struct writeback_control *wbc,
3756                         struct extent_page_data *epd)
3757 {
3758         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3759         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3760         u64 offset = eb->start;
3761         unsigned long i, num_pages;
3762         unsigned long bio_flags = 0;
3763         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3764         int ret = 0;
3765
3766         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3767         num_pages = num_extent_pages(eb->start, eb->len);
3768         atomic_set(&eb->io_pages, num_pages);
3769         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3770                 bio_flags = EXTENT_BIO_TREE_LOG;
3771
3772         for (i = 0; i < num_pages; i++) {
3773                 struct page *p = eb->pages[i];
3774
3775                 clear_page_dirty_for_io(p);
3776                 set_page_writeback(p);
3777                 ret = submit_extent_page(rw, tree, wbc, p, offset >> 9,
3778                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3779                                          -1, end_bio_extent_buffer_writepage,
3780                                          0, epd->bio_flags, bio_flags, false);
3781                 epd->bio_flags = bio_flags;
3782                 if (ret) {
3783                         set_btree_ioerr(p);
3784                         end_page_writeback(p);
3785                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3786                                 end_extent_buffer_writeback(eb);
3787                         ret = -EIO;
3788                         break;
3789                 }
3790                 offset += PAGE_CACHE_SIZE;
3791                 update_nr_written(p, wbc, 1);
3792                 unlock_page(p);
3793         }
3794
3795         if (unlikely(ret)) {
3796                 for (; i < num_pages; i++) {
3797                         struct page *p = eb->pages[i];
3798                         clear_page_dirty_for_io(p);
3799                         unlock_page(p);
3800                 }
3801         }
3802
3803         return ret;
3804 }
3805
3806 int btree_write_cache_pages(struct address_space *mapping,
3807                                    struct writeback_control *wbc)
3808 {
3809         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3810         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3811         struct extent_buffer *eb, *prev_eb = NULL;
3812         struct extent_page_data epd = {
3813                 .bio = NULL,
3814                 .tree = tree,
3815                 .extent_locked = 0,
3816                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3817                 .bio_flags = 0,
3818         };
3819         int ret = 0;
3820         int done = 0;
3821         int nr_to_write_done = 0;
3822         struct pagevec pvec;
3823         int nr_pages;
3824         pgoff_t index;
3825         pgoff_t end;            /* Inclusive */
3826         int scanned = 0;
3827         int tag;
3828
3829         pagevec_init(&pvec, 0);
3830         if (wbc->range_cyclic) {
3831                 index = mapping->writeback_index; /* Start from prev offset */
3832                 end = -1;
3833         } else {
3834                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3835                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3836                 scanned = 1;
3837         }
3838         if (wbc->sync_mode == WB_SYNC_ALL)
3839                 tag = PAGECACHE_TAG_TOWRITE;
3840         else
3841                 tag = PAGECACHE_TAG_DIRTY;
3842 retry:
3843         if (wbc->sync_mode == WB_SYNC_ALL)
3844                 tag_pages_for_writeback(mapping, index, end);
3845         while (!done && !nr_to_write_done && (index <= end) &&
3846                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3847                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3848                 unsigned i;
3849
3850                 scanned = 1;
3851                 for (i = 0; i < nr_pages; i++) {
3852                         struct page *page = pvec.pages[i];
3853
3854                         if (!PagePrivate(page))
3855                                 continue;
3856
3857                         if (!wbc->range_cyclic && page->index > end) {
3858                                 done = 1;
3859                                 break;
3860                         }
3861
3862                         spin_lock(&mapping->private_lock);
3863                         if (!PagePrivate(page)) {
3864                                 spin_unlock(&mapping->private_lock);
3865                                 continue;
3866                         }
3867
3868                         eb = (struct extent_buffer *)page->private;
3869
3870                         /*
3871                          * Shouldn't happen and normally this would be a BUG_ON
3872                          * but no sense in crashing the users box for something
3873                          * we can survive anyway.
3874                          */
3875                         if (WARN_ON(!eb)) {
3876                                 spin_unlock(&mapping->private_lock);
3877                                 continue;
3878                         }
3879
3880                         if (eb == prev_eb) {
3881                                 spin_unlock(&mapping->private_lock);
3882                                 continue;
3883                         }
3884
3885                         ret = atomic_inc_not_zero(&eb->refs);
3886                         spin_unlock(&mapping->private_lock);
3887                         if (!ret)
3888                                 continue;
3889
3890                         prev_eb = eb;
3891                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3892                         if (!ret) {
3893                                 free_extent_buffer(eb);
3894                                 continue;
3895                         }
3896
3897                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3898                         if (ret) {
3899                                 done = 1;
3900                                 free_extent_buffer(eb);
3901                                 break;
3902                         }
3903                         free_extent_buffer(eb);
3904
3905                         /*
3906                          * the filesystem may choose to bump up nr_to_write.
3907                          * We have to make sure to honor the new nr_to_write
3908                          * at any time
3909                          */
3910                         nr_to_write_done = wbc->nr_to_write <= 0;
3911                 }
3912                 pagevec_release(&pvec);
3913                 cond_resched();
3914         }
3915         if (!scanned && !done) {
3916                 /*
3917                  * We hit the last page and there is more work to be done: wrap
3918                  * back to the start of the file
3919                  */
3920                 scanned = 1;
3921                 index = 0;
3922                 goto retry;
3923         }
3924         flush_write_bio(&epd);
3925         return ret;
3926 }
3927
3928 /**
3929  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3930  * @mapping: address space structure to write
3931  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3932  * @writepage: function called for each page
3933  * @data: data passed to writepage function
3934  *
3935  * If a page is already under I/O, write_cache_pages() skips it, even
3936  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3937  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3938  * and msync() need to guarantee that all the data which was dirty at the time
3939  * the call was made get new I/O started against them.  If wbc->sync_mode is
3940  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3941  * existing IO to complete.
3942  */
3943 static int extent_write_cache_pages(struct extent_io_tree *tree,
3944                              struct address_space *mapping,
3945                              struct writeback_control *wbc,
3946                              writepage_t writepage, void *data,
3947                              void (*flush_fn)(void *))
3948 {
3949         struct inode *inode = mapping->host;
3950         int ret = 0;
3951         int done = 0;
3952         int err = 0;
3953         int nr_to_write_done = 0;
3954         struct pagevec pvec;
3955         int nr_pages;
3956         pgoff_t index;
3957         pgoff_t end;            /* Inclusive */
3958         int scanned = 0;
3959         int tag;
3960
3961         /*
3962          * We have to hold onto the inode so that ordered extents can do their
3963          * work when the IO finishes.  The alternative to this is failing to add
3964          * an ordered extent if the igrab() fails there and that is a huge pain
3965          * to deal with, so instead just hold onto the inode throughout the
3966          * writepages operation.  If it fails here we are freeing up the inode
3967          * anyway and we'd rather not waste our time writing out stuff that is
3968          * going to be truncated anyway.
3969          */
3970         if (!igrab(inode))
3971                 return 0;
3972
3973         pagevec_init(&pvec, 0);
3974         if (wbc->range_cyclic) {
3975                 index = mapping->writeback_index; /* Start from prev offset */
3976                 end = -1;
3977         } else {
3978                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3979                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3980                 scanned = 1;
3981         }
3982         if (wbc->sync_mode == WB_SYNC_ALL)
3983                 tag = PAGECACHE_TAG_TOWRITE;
3984         else
3985                 tag = PAGECACHE_TAG_DIRTY;
3986 retry:
3987         if (wbc->sync_mode == WB_SYNC_ALL)
3988                 tag_pages_for_writeback(mapping, index, end);
3989         while (!done && !nr_to_write_done && (index <= end) &&
3990                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3991                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3992                 unsigned i;
3993
3994                 scanned = 1;
3995                 for (i = 0; i < nr_pages; i++) {
3996                         struct page *page = pvec.pages[i];
3997
3998                         /*
3999                          * At this point we hold neither mapping->tree_lock nor
4000                          * lock on the page itself: the page may be truncated or
4001                          * invalidated (changing page->mapping to NULL), or even
4002                          * swizzled back from swapper_space to tmpfs file
4003                          * mapping
4004                          */
4005                         if (!trylock_page(page)) {
4006                                 flush_fn(data);
4007                                 lock_page(page);
4008                         }
4009
4010                         if (unlikely(page->mapping != mapping)) {
4011                                 unlock_page(page);
4012                                 continue;
4013                         }
4014
4015                         if (!wbc->range_cyclic && page->index > end) {
4016                                 done = 1;
4017                                 unlock_page(page);
4018                                 continue;
4019                         }
4020
4021                         if (wbc->sync_mode != WB_SYNC_NONE) {
4022                                 if (PageWriteback(page))
4023                                         flush_fn(data);
4024                                 wait_on_page_writeback(page);
4025                         }
4026
4027                         if (PageWriteback(page) ||
4028                             !clear_page_dirty_for_io(page)) {
4029                                 unlock_page(page);
4030                                 continue;
4031                         }
4032
4033                         ret = (*writepage)(page, wbc, data);
4034
4035                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4036                                 unlock_page(page);
4037                                 ret = 0;
4038                         }
4039                         if (!err && ret < 0)
4040                                 err = ret;
4041
4042                         /*
4043                          * the filesystem may choose to bump up nr_to_write.
4044                          * We have to make sure to honor the new nr_to_write
4045                          * at any time
4046                          */
4047                         nr_to_write_done = wbc->nr_to_write <= 0;
4048                 }
4049                 pagevec_release(&pvec);
4050                 cond_resched();
4051         }
4052         if (!scanned && !done && !err) {
4053                 /*
4054                  * We hit the last page and there is more work to be done: wrap
4055                  * back to the start of the file
4056                  */
4057                 scanned = 1;
4058                 index = 0;
4059                 goto retry;
4060         }
4061         btrfs_add_delayed_iput(inode);
4062         return err;
4063 }
4064
4065 static void flush_epd_write_bio(struct extent_page_data *epd)
4066 {
4067         if (epd->bio) {
4068                 int rw = WRITE;
4069                 int ret;
4070
4071                 if (epd->sync_io)
4072                         rw = WRITE_SYNC;
4073
4074                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
4075                 BUG_ON(ret < 0); /* -ENOMEM */
4076                 epd->bio = NULL;
4077         }
4078 }
4079
4080 static noinline void flush_write_bio(void *data)
4081 {
4082         struct extent_page_data *epd = data;
4083         flush_epd_write_bio(epd);
4084 }
4085
4086 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4087                           get_extent_t *get_extent,
4088                           struct writeback_control *wbc)
4089 {
4090         int ret;
4091         struct extent_page_data epd = {
4092                 .bio = NULL,
4093                 .tree = tree,
4094                 .get_extent = get_extent,
4095                 .extent_locked = 0,
4096                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4097                 .bio_flags = 0,
4098         };
4099
4100         ret = __extent_writepage(page, wbc, &epd);
4101
4102         flush_epd_write_bio(&epd);
4103         return ret;
4104 }
4105
4106 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4107                               u64 start, u64 end, get_extent_t *get_extent,
4108                               int mode)
4109 {
4110         int ret = 0;
4111         struct address_space *mapping = inode->i_mapping;
4112         struct page *page;
4113         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
4114                 PAGE_CACHE_SHIFT;
4115
4116         struct extent_page_data epd = {
4117                 .bio = NULL,
4118                 .tree = tree,
4119                 .get_extent = get_extent,
4120                 .extent_locked = 1,
4121                 .sync_io = mode == WB_SYNC_ALL,
4122                 .bio_flags = 0,
4123         };
4124         struct writeback_control wbc_writepages = {
4125                 .sync_mode      = mode,
4126                 .nr_to_write    = nr_pages * 2,
4127                 .range_start    = start,
4128                 .range_end      = end + 1,
4129         };
4130
4131         while (start <= end) {
4132                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
4133                 if (clear_page_dirty_for_io(page))
4134                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4135                 else {
4136                         if (tree->ops && tree->ops->writepage_end_io_hook)
4137                                 tree->ops->writepage_end_io_hook(page, start,
4138                                                  start + PAGE_CACHE_SIZE - 1,
4139                                                  NULL, 1);
4140                         unlock_page(page);
4141                 }
4142                 page_cache_release(page);
4143                 start += PAGE_CACHE_SIZE;
4144         }
4145
4146         flush_epd_write_bio(&epd);
4147         return ret;
4148 }
4149
4150 int extent_writepages(struct extent_io_tree *tree,
4151                       struct address_space *mapping,
4152                       get_extent_t *get_extent,
4153                       struct writeback_control *wbc)
4154 {
4155         int ret = 0;
4156         struct extent_page_data epd = {
4157                 .bio = NULL,
4158                 .tree = tree,
4159                 .get_extent = get_extent,
4160                 .extent_locked = 0,
4161                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4162                 .bio_flags = 0,
4163         };
4164
4165         ret = extent_write_cache_pages(tree, mapping, wbc,
4166                                        __extent_writepage, &epd,
4167                                        flush_write_bio);
4168         flush_epd_write_bio(&epd);
4169         return ret;
4170 }
4171
4172 int extent_readpages(struct extent_io_tree *tree,
4173                      struct address_space *mapping,
4174                      struct list_head *pages, unsigned nr_pages,
4175                      get_extent_t get_extent)
4176 {
4177         struct bio *bio = NULL;
4178         unsigned page_idx;
4179         unsigned long bio_flags = 0;
4180         struct page *pagepool[16];
4181         struct page *page;
4182         struct extent_map *em_cached = NULL;
4183         int nr = 0;
4184         u64 prev_em_start = (u64)-1;
4185
4186         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4187                 page = list_entry(pages->prev, struct page, lru);
4188
4189                 prefetchw(&page->flags);
4190                 list_del(&page->lru);
4191                 if (add_to_page_cache_lru(page, mapping,
4192                                         page->index, GFP_NOFS)) {
4193                         page_cache_release(page);
4194                         continue;
4195                 }
4196
4197                 pagepool[nr++] = page;
4198                 if (nr < ARRAY_SIZE(pagepool))
4199                         continue;
4200                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4201                                    &bio, 0, &bio_flags, READ, &prev_em_start);
4202                 nr = 0;
4203         }
4204         if (nr)
4205                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4206                                    &bio, 0, &bio_flags, READ, &prev_em_start);
4207
4208         if (em_cached)
4209                 free_extent_map(em_cached);
4210
4211         BUG_ON(!list_empty(pages));
4212         if (bio)
4213                 return submit_one_bio(READ, bio, 0, bio_flags);
4214         return 0;
4215 }
4216
4217 /*
4218  * basic invalidatepage code, this waits on any locked or writeback
4219  * ranges corresponding to the page, and then deletes any extent state
4220  * records from the tree
4221  */
4222 int extent_invalidatepage(struct extent_io_tree *tree,
4223                           struct page *page, unsigned long offset)
4224 {
4225         struct extent_state *cached_state = NULL;
4226         u64 start = page_offset(page);
4227         u64 end = start + PAGE_CACHE_SIZE - 1;
4228         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4229
4230         start += ALIGN(offset, blocksize);
4231         if (start > end)
4232                 return 0;
4233
4234         lock_extent_bits(tree, start, end, &cached_state);
4235         wait_on_page_writeback(page);
4236         clear_extent_bit(tree, start, end,
4237                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4238                          EXTENT_DO_ACCOUNTING,
4239                          1, 1, &cached_state, GFP_NOFS);
4240         return 0;
4241 }
4242
4243 /*
4244  * a helper for releasepage, this tests for areas of the page that
4245  * are locked or under IO and drops the related state bits if it is safe
4246  * to drop the page.
4247  */
4248 static int try_release_extent_state(struct extent_map_tree *map,
4249                                     struct extent_io_tree *tree,
4250                                     struct page *page, gfp_t mask)
4251 {
4252         u64 start = page_offset(page);
4253         u64 end = start + PAGE_CACHE_SIZE - 1;
4254         int ret = 1;
4255
4256         if (test_range_bit(tree, start, end,
4257                            EXTENT_IOBITS, 0, NULL))
4258                 ret = 0;
4259         else {
4260                 if ((mask & GFP_NOFS) == GFP_NOFS)
4261                         mask = GFP_NOFS;
4262                 /*
4263                  * at this point we can safely clear everything except the
4264                  * locked bit and the nodatasum bit
4265                  */
4266                 ret = clear_extent_bit(tree, start, end,
4267                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4268                                  0, 0, NULL, mask);
4269
4270                 /* if clear_extent_bit failed for enomem reasons,
4271                  * we can't allow the release to continue.
4272                  */
4273                 if (ret < 0)
4274                         ret = 0;
4275                 else
4276                         ret = 1;
4277         }
4278         return ret;
4279 }
4280
4281 /*
4282  * a helper for releasepage.  As long as there are no locked extents
4283  * in the range corresponding to the page, both state records and extent
4284  * map records are removed
4285  */
4286 int try_release_extent_mapping(struct extent_map_tree *map,
4287                                struct extent_io_tree *tree, struct page *page,
4288                                gfp_t mask)
4289 {
4290         struct extent_map *em;
4291         u64 start = page_offset(page);
4292         u64 end = start + PAGE_CACHE_SIZE - 1;
4293
4294         if (gfpflags_allow_blocking(mask) &&
4295             page->mapping->host->i_size > SZ_16M) {
4296                 u64 len;
4297                 while (start <= end) {
4298                         len = end - start + 1;
4299                         write_lock(&map->lock);
4300                         em = lookup_extent_mapping(map, start, len);
4301                         if (!em) {
4302                                 write_unlock(&map->lock);
4303                                 break;
4304                         }
4305                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4306                             em->start != start) {
4307                                 write_unlock(&map->lock);
4308                                 free_extent_map(em);
4309                                 break;
4310                         }
4311                         if (!test_range_bit(tree, em->start,
4312                                             extent_map_end(em) - 1,
4313                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4314                                             0, NULL)) {
4315                                 remove_extent_mapping(map, em);
4316                                 /* once for the rb tree */
4317                                 free_extent_map(em);
4318                         }
4319                         start = extent_map_end(em);
4320                         write_unlock(&map->lock);
4321
4322                         /* once for us */
4323                         free_extent_map(em);
4324                 }
4325         }
4326         return try_release_extent_state(map, tree, page, mask);
4327 }
4328
4329 /*
4330  * helper function for fiemap, which doesn't want to see any holes.
4331  * This maps until we find something past 'last'
4332  */
4333 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4334                                                 u64 offset,
4335                                                 u64 last,
4336                                                 get_extent_t *get_extent)
4337 {
4338         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4339         struct extent_map *em;
4340         u64 len;
4341
4342         if (offset >= last)
4343                 return NULL;
4344
4345         while (1) {
4346                 len = last - offset;
4347                 if (len == 0)
4348                         break;
4349                 len = ALIGN(len, sectorsize);
4350                 em = get_extent(inode, NULL, 0, offset, len, 0);
4351                 if (IS_ERR_OR_NULL(em))
4352                         return em;
4353
4354                 /* if this isn't a hole return it */
4355                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4356                     em->block_start != EXTENT_MAP_HOLE) {
4357                         return em;
4358                 }
4359
4360                 /* this is a hole, advance to the next extent */
4361                 offset = extent_map_end(em);
4362                 free_extent_map(em);
4363                 if (offset >= last)
4364                         break;
4365         }
4366         return NULL;
4367 }
4368
4369 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4370                 __u64 start, __u64 len, get_extent_t *get_extent)
4371 {
4372         int ret = 0;
4373         u64 off = start;
4374         u64 max = start + len;
4375         u32 flags = 0;
4376         u32 found_type;
4377         u64 last;
4378         u64 last_for_get_extent = 0;
4379         u64 disko = 0;
4380         u64 isize = i_size_read(inode);
4381         struct btrfs_key found_key;
4382         struct extent_map *em = NULL;
4383         struct extent_state *cached_state = NULL;
4384         struct btrfs_path *path;
4385         struct btrfs_root *root = BTRFS_I(inode)->root;
4386         int end = 0;
4387         u64 em_start = 0;
4388         u64 em_len = 0;
4389         u64 em_end = 0;
4390
4391         if (len == 0)
4392                 return -EINVAL;
4393
4394         path = btrfs_alloc_path();
4395         if (!path)
4396                 return -ENOMEM;
4397         path->leave_spinning = 1;
4398
4399         start = round_down(start, BTRFS_I(inode)->root->sectorsize);
4400         len = round_up(max, BTRFS_I(inode)->root->sectorsize) - start;
4401
4402         /*
4403          * lookup the last file extent.  We're not using i_size here
4404          * because there might be preallocation past i_size
4405          */
4406         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode), -1,
4407                                        0);
4408         if (ret < 0) {
4409                 btrfs_free_path(path);
4410                 return ret;
4411         }
4412         WARN_ON(!ret);
4413         path->slots[0]--;
4414         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4415         found_type = found_key.type;
4416
4417         /* No extents, but there might be delalloc bits */
4418         if (found_key.objectid != btrfs_ino(inode) ||
4419             found_type != BTRFS_EXTENT_DATA_KEY) {
4420                 /* have to trust i_size as the end */
4421                 last = (u64)-1;
4422                 last_for_get_extent = isize;
4423         } else {
4424                 /*
4425                  * remember the start of the last extent.  There are a
4426                  * bunch of different factors that go into the length of the
4427                  * extent, so its much less complex to remember where it started
4428                  */
4429                 last = found_key.offset;
4430                 last_for_get_extent = last + 1;
4431         }
4432         btrfs_release_path(path);
4433
4434         /*
4435          * we might have some extents allocated but more delalloc past those
4436          * extents.  so, we trust isize unless the start of the last extent is
4437          * beyond isize
4438          */
4439         if (last < isize) {
4440                 last = (u64)-1;
4441                 last_for_get_extent = isize;
4442         }
4443
4444         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4445                          &cached_state);
4446
4447         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4448                                    get_extent);
4449         if (!em)
4450                 goto out;
4451         if (IS_ERR(em)) {
4452                 ret = PTR_ERR(em);
4453                 goto out;
4454         }
4455
4456         while (!end) {
4457                 u64 offset_in_extent = 0;
4458
4459                 /* break if the extent we found is outside the range */
4460                 if (em->start >= max || extent_map_end(em) < off)
4461                         break;
4462
4463                 /*
4464                  * get_extent may return an extent that starts before our
4465                  * requested range.  We have to make sure the ranges
4466                  * we return to fiemap always move forward and don't
4467                  * overlap, so adjust the offsets here
4468                  */
4469                 em_start = max(em->start, off);
4470
4471                 /*
4472                  * record the offset from the start of the extent
4473                  * for adjusting the disk offset below.  Only do this if the
4474                  * extent isn't compressed since our in ram offset may be past
4475                  * what we have actually allocated on disk.
4476                  */
4477                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4478                         offset_in_extent = em_start - em->start;
4479                 em_end = extent_map_end(em);
4480                 em_len = em_end - em_start;
4481                 disko = 0;
4482                 flags = 0;
4483
4484                 /*
4485                  * bump off for our next call to get_extent
4486                  */
4487                 off = extent_map_end(em);
4488                 if (off >= max)
4489                         end = 1;
4490
4491                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4492                         end = 1;
4493                         flags |= FIEMAP_EXTENT_LAST;
4494                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4495                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4496                                   FIEMAP_EXTENT_NOT_ALIGNED);
4497                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4498                         flags |= (FIEMAP_EXTENT_DELALLOC |
4499                                   FIEMAP_EXTENT_UNKNOWN);
4500                 } else if (fieinfo->fi_extents_max) {
4501                         u64 bytenr = em->block_start -
4502                                 (em->start - em->orig_start);
4503
4504                         disko = em->block_start + offset_in_extent;
4505
4506                         /*
4507                          * As btrfs supports shared space, this information
4508                          * can be exported to userspace tools via
4509                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4510                          * then we're just getting a count and we can skip the
4511                          * lookup stuff.
4512                          */
4513                         ret = btrfs_check_shared(NULL, root->fs_info,
4514                                                  root->objectid,
4515                                                  btrfs_ino(inode), bytenr);
4516                         if (ret < 0)
4517                                 goto out_free;
4518                         if (ret)
4519                                 flags |= FIEMAP_EXTENT_SHARED;
4520                         ret = 0;
4521                 }
4522                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4523                         flags |= FIEMAP_EXTENT_ENCODED;
4524                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4525                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4526
4527                 free_extent_map(em);
4528                 em = NULL;
4529                 if ((em_start >= last) || em_len == (u64)-1 ||
4530                    (last == (u64)-1 && isize <= em_end)) {
4531                         flags |= FIEMAP_EXTENT_LAST;
4532                         end = 1;
4533                 }
4534
4535                 /* now scan forward to see if this is really the last extent. */
4536                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4537                                            get_extent);
4538                 if (IS_ERR(em)) {
4539                         ret = PTR_ERR(em);
4540                         goto out;
4541                 }
4542                 if (!em) {
4543                         flags |= FIEMAP_EXTENT_LAST;
4544                         end = 1;
4545                 }
4546                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4547                                               em_len, flags);
4548                 if (ret) {
4549                         if (ret == 1)
4550                                 ret = 0;
4551                         goto out_free;
4552                 }
4553         }
4554 out_free:
4555         free_extent_map(em);
4556 out:
4557         btrfs_free_path(path);
4558         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4559                              &cached_state, GFP_NOFS);
4560         return ret;
4561 }
4562
4563 static void __free_extent_buffer(struct extent_buffer *eb)
4564 {
4565         btrfs_leak_debug_del(&eb->leak_list);
4566         kmem_cache_free(extent_buffer_cache, eb);
4567 }
4568
4569 int extent_buffer_under_io(struct extent_buffer *eb)
4570 {
4571         return (atomic_read(&eb->io_pages) ||
4572                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4573                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4574 }
4575
4576 /*
4577  * Helper for releasing extent buffer page.
4578  */
4579 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4580 {
4581         unsigned long index;
4582         struct page *page;
4583         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4584
4585         BUG_ON(extent_buffer_under_io(eb));
4586
4587         index = num_extent_pages(eb->start, eb->len);
4588         if (index == 0)
4589                 return;
4590
4591         do {
4592                 index--;
4593                 page = eb->pages[index];
4594                 if (!page)
4595                         continue;
4596                 if (mapped)
4597                         spin_lock(&page->mapping->private_lock);
4598                 /*
4599                  * We do this since we'll remove the pages after we've
4600                  * removed the eb from the radix tree, so we could race
4601                  * and have this page now attached to the new eb.  So
4602                  * only clear page_private if it's still connected to
4603                  * this eb.
4604                  */
4605                 if (PagePrivate(page) &&
4606                     page->private == (unsigned long)eb) {
4607                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4608                         BUG_ON(PageDirty(page));
4609                         BUG_ON(PageWriteback(page));
4610                         /*
4611                          * We need to make sure we haven't be attached
4612                          * to a new eb.
4613                          */
4614                         ClearPagePrivate(page);
4615                         set_page_private(page, 0);
4616                         /* One for the page private */
4617                         page_cache_release(page);
4618                 }
4619
4620                 if (mapped)
4621                         spin_unlock(&page->mapping->private_lock);
4622
4623                 /* One for when we alloced the page */
4624                 page_cache_release(page);
4625         } while (index != 0);
4626 }
4627
4628 /*
4629  * Helper for releasing the extent buffer.
4630  */
4631 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4632 {
4633         btrfs_release_extent_buffer_page(eb);
4634         __free_extent_buffer(eb);
4635 }
4636
4637 static struct extent_buffer *
4638 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4639                       unsigned long len)
4640 {
4641         struct extent_buffer *eb = NULL;
4642
4643         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4644         eb->start = start;
4645         eb->len = len;
4646         eb->fs_info = fs_info;
4647         eb->bflags = 0;
4648         rwlock_init(&eb->lock);
4649         atomic_set(&eb->write_locks, 0);
4650         atomic_set(&eb->read_locks, 0);
4651         atomic_set(&eb->blocking_readers, 0);
4652         atomic_set(&eb->blocking_writers, 0);
4653         atomic_set(&eb->spinning_readers, 0);
4654         atomic_set(&eb->spinning_writers, 0);
4655         eb->lock_nested = 0;
4656         init_waitqueue_head(&eb->write_lock_wq);
4657         init_waitqueue_head(&eb->read_lock_wq);
4658
4659         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4660
4661         spin_lock_init(&eb->refs_lock);
4662         atomic_set(&eb->refs, 1);
4663         atomic_set(&eb->io_pages, 0);
4664
4665         /*
4666          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4667          */
4668         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4669                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4670         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4671
4672         return eb;
4673 }
4674
4675 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4676 {
4677         unsigned long i;
4678         struct page *p;
4679         struct extent_buffer *new;
4680         unsigned long num_pages = num_extent_pages(src->start, src->len);
4681
4682         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4683         if (new == NULL)
4684                 return NULL;
4685
4686         for (i = 0; i < num_pages; i++) {
4687                 p = alloc_page(GFP_NOFS);
4688                 if (!p) {
4689                         btrfs_release_extent_buffer(new);
4690                         return NULL;
4691                 }
4692                 attach_extent_buffer_page(new, p);
4693                 WARN_ON(PageDirty(p));
4694                 SetPageUptodate(p);
4695                 new->pages[i] = p;
4696         }
4697
4698         copy_extent_buffer(new, src, 0, 0, src->len);
4699         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4700         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4701
4702         return new;
4703 }
4704
4705 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4706                                                   u64 start, unsigned long len)
4707 {
4708         struct extent_buffer *eb;
4709         unsigned long num_pages;
4710         unsigned long i;
4711
4712         num_pages = num_extent_pages(start, len);
4713
4714         eb = __alloc_extent_buffer(fs_info, start, len);
4715         if (!eb)
4716                 return NULL;
4717
4718         for (i = 0; i < num_pages; i++) {
4719                 eb->pages[i] = alloc_page(GFP_NOFS);
4720                 if (!eb->pages[i])
4721                         goto err;
4722         }
4723         set_extent_buffer_uptodate(eb);
4724         btrfs_set_header_nritems(eb, 0);
4725         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4726
4727         return eb;
4728 err:
4729         for (; i > 0; i--)
4730                 __free_page(eb->pages[i - 1]);
4731         __free_extent_buffer(eb);
4732         return NULL;
4733 }
4734
4735 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4736                                                 u64 start)
4737 {
4738         unsigned long len;
4739
4740         if (!fs_info) {
4741                 /*
4742                  * Called only from tests that don't always have a fs_info
4743                  * available, but we know that nodesize is 4096
4744                  */
4745                 len = 4096;
4746         } else {
4747                 len = fs_info->tree_root->nodesize;
4748         }
4749
4750         return __alloc_dummy_extent_buffer(fs_info, start, len);
4751 }
4752
4753 static void check_buffer_tree_ref(struct extent_buffer *eb)
4754 {
4755         int refs;
4756         /* the ref bit is tricky.  We have to make sure it is set
4757          * if we have the buffer dirty.   Otherwise the
4758          * code to free a buffer can end up dropping a dirty
4759          * page
4760          *
4761          * Once the ref bit is set, it won't go away while the
4762          * buffer is dirty or in writeback, and it also won't
4763          * go away while we have the reference count on the
4764          * eb bumped.
4765          *
4766          * We can't just set the ref bit without bumping the
4767          * ref on the eb because free_extent_buffer might
4768          * see the ref bit and try to clear it.  If this happens
4769          * free_extent_buffer might end up dropping our original
4770          * ref by mistake and freeing the page before we are able
4771          * to add one more ref.
4772          *
4773          * So bump the ref count first, then set the bit.  If someone
4774          * beat us to it, drop the ref we added.
4775          */
4776         refs = atomic_read(&eb->refs);
4777         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4778                 return;
4779
4780         spin_lock(&eb->refs_lock);
4781         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4782                 atomic_inc(&eb->refs);
4783         spin_unlock(&eb->refs_lock);
4784 }
4785
4786 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4787                 struct page *accessed)
4788 {
4789         unsigned long num_pages, i;
4790
4791         check_buffer_tree_ref(eb);
4792
4793         num_pages = num_extent_pages(eb->start, eb->len);
4794         for (i = 0; i < num_pages; i++) {
4795                 struct page *p = eb->pages[i];
4796
4797                 if (p != accessed)
4798                         mark_page_accessed(p);
4799         }
4800 }
4801
4802 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4803                                          u64 start)
4804 {
4805         struct extent_buffer *eb;
4806
4807         rcu_read_lock();
4808         eb = radix_tree_lookup(&fs_info->buffer_radix,
4809                                start >> PAGE_CACHE_SHIFT);
4810         if (eb && atomic_inc_not_zero(&eb->refs)) {
4811                 rcu_read_unlock();
4812                 /*
4813                  * Lock our eb's refs_lock to avoid races with
4814                  * free_extent_buffer. When we get our eb it might be flagged
4815                  * with EXTENT_BUFFER_STALE and another task running
4816                  * free_extent_buffer might have seen that flag set,
4817                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4818                  * writeback flags not set) and it's still in the tree (flag
4819                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4820                  * of decrementing the extent buffer's reference count twice.
4821                  * So here we could race and increment the eb's reference count,
4822                  * clear its stale flag, mark it as dirty and drop our reference
4823                  * before the other task finishes executing free_extent_buffer,
4824                  * which would later result in an attempt to free an extent
4825                  * buffer that is dirty.
4826                  */
4827                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4828                         spin_lock(&eb->refs_lock);
4829                         spin_unlock(&eb->refs_lock);
4830                 }
4831                 mark_extent_buffer_accessed(eb, NULL);
4832                 return eb;
4833         }
4834         rcu_read_unlock();
4835
4836         return NULL;
4837 }
4838
4839 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4840 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4841                                                u64 start)
4842 {
4843         struct extent_buffer *eb, *exists = NULL;
4844         int ret;
4845
4846         eb = find_extent_buffer(fs_info, start);
4847         if (eb)
4848                 return eb;
4849         eb = alloc_dummy_extent_buffer(fs_info, start);
4850         if (!eb)
4851                 return NULL;
4852         eb->fs_info = fs_info;
4853 again:
4854         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4855         if (ret)
4856                 goto free_eb;
4857         spin_lock(&fs_info->buffer_lock);
4858         ret = radix_tree_insert(&fs_info->buffer_radix,
4859                                 start >> PAGE_CACHE_SHIFT, eb);
4860         spin_unlock(&fs_info->buffer_lock);
4861         radix_tree_preload_end();
4862         if (ret == -EEXIST) {
4863                 exists = find_extent_buffer(fs_info, start);
4864                 if (exists)
4865                         goto free_eb;
4866                 else
4867                         goto again;
4868         }
4869         check_buffer_tree_ref(eb);
4870         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4871
4872         /*
4873          * We will free dummy extent buffer's if they come into
4874          * free_extent_buffer with a ref count of 2, but if we are using this we
4875          * want the buffers to stay in memory until we're done with them, so
4876          * bump the ref count again.
4877          */
4878         atomic_inc(&eb->refs);
4879         return eb;
4880 free_eb:
4881         btrfs_release_extent_buffer(eb);
4882         return exists;
4883 }
4884 #endif
4885
4886 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4887                                           u64 start)
4888 {
4889         unsigned long len = fs_info->tree_root->nodesize;
4890         unsigned long num_pages = num_extent_pages(start, len);
4891         unsigned long i;
4892         unsigned long index = start >> PAGE_CACHE_SHIFT;
4893         struct extent_buffer *eb;
4894         struct extent_buffer *exists = NULL;
4895         struct page *p;
4896         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4897         int uptodate = 1;
4898         int ret;
4899
4900         eb = find_extent_buffer(fs_info, start);
4901         if (eb)
4902                 return eb;
4903
4904         eb = __alloc_extent_buffer(fs_info, start, len);
4905         if (!eb)
4906                 return NULL;
4907
4908         for (i = 0; i < num_pages; i++, index++) {
4909                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4910                 if (!p)
4911                         goto free_eb;
4912
4913                 spin_lock(&mapping->private_lock);
4914                 if (PagePrivate(p)) {
4915                         /*
4916                          * We could have already allocated an eb for this page
4917                          * and attached one so lets see if we can get a ref on
4918                          * the existing eb, and if we can we know it's good and
4919                          * we can just return that one, else we know we can just
4920                          * overwrite page->private.
4921                          */
4922                         exists = (struct extent_buffer *)p->private;
4923                         if (atomic_inc_not_zero(&exists->refs)) {
4924                                 spin_unlock(&mapping->private_lock);
4925                                 unlock_page(p);
4926                                 page_cache_release(p);
4927                                 mark_extent_buffer_accessed(exists, p);
4928                                 goto free_eb;
4929                         }
4930                         exists = NULL;
4931
4932                         /*
4933                          * Do this so attach doesn't complain and we need to
4934                          * drop the ref the old guy had.
4935                          */
4936                         ClearPagePrivate(p);
4937                         WARN_ON(PageDirty(p));
4938                         page_cache_release(p);
4939                 }
4940                 attach_extent_buffer_page(eb, p);
4941                 spin_unlock(&mapping->private_lock);
4942                 WARN_ON(PageDirty(p));
4943                 eb->pages[i] = p;
4944                 if (!PageUptodate(p))
4945                         uptodate = 0;
4946
4947                 /*
4948                  * see below about how we avoid a nasty race with release page
4949                  * and why we unlock later
4950                  */
4951         }
4952         if (uptodate)
4953                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4954 again:
4955         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4956         if (ret)
4957                 goto free_eb;
4958
4959         spin_lock(&fs_info->buffer_lock);
4960         ret = radix_tree_insert(&fs_info->buffer_radix,
4961                                 start >> PAGE_CACHE_SHIFT, eb);
4962         spin_unlock(&fs_info->buffer_lock);
4963         radix_tree_preload_end();
4964         if (ret == -EEXIST) {
4965                 exists = find_extent_buffer(fs_info, start);
4966                 if (exists)
4967                         goto free_eb;
4968                 else
4969                         goto again;
4970         }
4971         /* add one reference for the tree */
4972         check_buffer_tree_ref(eb);
4973         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4974
4975         /*
4976          * there is a race where release page may have
4977          * tried to find this extent buffer in the radix
4978          * but failed.  It will tell the VM it is safe to
4979          * reclaim the, and it will clear the page private bit.
4980          * We must make sure to set the page private bit properly
4981          * after the extent buffer is in the radix tree so
4982          * it doesn't get lost
4983          */
4984         SetPageChecked(eb->pages[0]);
4985         for (i = 1; i < num_pages; i++) {
4986                 p = eb->pages[i];
4987                 ClearPageChecked(p);
4988                 unlock_page(p);
4989         }
4990         unlock_page(eb->pages[0]);
4991         return eb;
4992
4993 free_eb:
4994         WARN_ON(!atomic_dec_and_test(&eb->refs));
4995         for (i = 0; i < num_pages; i++) {
4996                 if (eb->pages[i])
4997                         unlock_page(eb->pages[i]);
4998         }
4999
5000         btrfs_release_extent_buffer(eb);
5001         return exists;
5002 }
5003
5004 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5005 {
5006         struct extent_buffer *eb =
5007                         container_of(head, struct extent_buffer, rcu_head);
5008
5009         __free_extent_buffer(eb);
5010 }
5011
5012 /* Expects to have eb->eb_lock already held */
5013 static int release_extent_buffer(struct extent_buffer *eb)
5014 {
5015         WARN_ON(atomic_read(&eb->refs) == 0);
5016         if (atomic_dec_and_test(&eb->refs)) {
5017                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5018                         struct btrfs_fs_info *fs_info = eb->fs_info;
5019
5020                         spin_unlock(&eb->refs_lock);
5021
5022                         spin_lock(&fs_info->buffer_lock);
5023                         radix_tree_delete(&fs_info->buffer_radix,
5024                                           eb->start >> PAGE_CACHE_SHIFT);
5025                         spin_unlock(&fs_info->buffer_lock);
5026                 } else {
5027                         spin_unlock(&eb->refs_lock);
5028                 }
5029
5030                 /* Should be safe to release our pages at this point */
5031                 btrfs_release_extent_buffer_page(eb);
5032 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5033                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5034                         __free_extent_buffer(eb);
5035                         return 1;
5036                 }
5037 #endif
5038                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5039                 return 1;
5040         }
5041         spin_unlock(&eb->refs_lock);
5042
5043         return 0;
5044 }
5045
5046 void free_extent_buffer(struct extent_buffer *eb)
5047 {
5048         int refs;
5049         int old;
5050         if (!eb)
5051                 return;
5052
5053         while (1) {
5054                 refs = atomic_read(&eb->refs);
5055                 if (refs <= 3)
5056                         break;
5057                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5058                 if (old == refs)
5059                         return;
5060         }
5061
5062         spin_lock(&eb->refs_lock);
5063         if (atomic_read(&eb->refs) == 2 &&
5064             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5065                 atomic_dec(&eb->refs);
5066
5067         if (atomic_read(&eb->refs) == 2 &&
5068             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5069             !extent_buffer_under_io(eb) &&
5070             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5071                 atomic_dec(&eb->refs);
5072
5073         /*
5074          * I know this is terrible, but it's temporary until we stop tracking
5075          * the uptodate bits and such for the extent buffers.
5076          */
5077         release_extent_buffer(eb);
5078 }
5079
5080 void free_extent_buffer_stale(struct extent_buffer *eb)
5081 {
5082         if (!eb)
5083                 return;
5084
5085         spin_lock(&eb->refs_lock);
5086         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5087
5088         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5089             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5090                 atomic_dec(&eb->refs);
5091         release_extent_buffer(eb);
5092 }
5093
5094 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5095 {
5096         unsigned long i;
5097         unsigned long num_pages;
5098         struct page *page;
5099
5100         num_pages = num_extent_pages(eb->start, eb->len);
5101
5102         for (i = 0; i < num_pages; i++) {
5103                 page = eb->pages[i];
5104                 if (!PageDirty(page))
5105                         continue;
5106
5107                 lock_page(page);
5108                 WARN_ON(!PagePrivate(page));
5109
5110                 clear_page_dirty_for_io(page);
5111                 spin_lock_irq(&page->mapping->tree_lock);
5112                 if (!PageDirty(page)) {
5113                         radix_tree_tag_clear(&page->mapping->page_tree,
5114                                                 page_index(page),
5115                                                 PAGECACHE_TAG_DIRTY);
5116                 }
5117                 spin_unlock_irq(&page->mapping->tree_lock);
5118                 ClearPageError(page);
5119                 unlock_page(page);
5120         }
5121         WARN_ON(atomic_read(&eb->refs) == 0);
5122 }
5123
5124 int set_extent_buffer_dirty(struct extent_buffer *eb)
5125 {
5126         unsigned long i;
5127         unsigned long num_pages;
5128         int was_dirty = 0;
5129
5130         check_buffer_tree_ref(eb);
5131
5132         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5133
5134         num_pages = num_extent_pages(eb->start, eb->len);
5135         WARN_ON(atomic_read(&eb->refs) == 0);
5136         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5137
5138         for (i = 0; i < num_pages; i++)
5139                 set_page_dirty(eb->pages[i]);
5140         return was_dirty;
5141 }
5142
5143 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5144 {
5145         unsigned long i;
5146         struct page *page;
5147         unsigned long num_pages;
5148
5149         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5150         num_pages = num_extent_pages(eb->start, eb->len);
5151         for (i = 0; i < num_pages; i++) {
5152                 page = eb->pages[i];
5153                 if (page)
5154                         ClearPageUptodate(page);
5155         }
5156 }
5157
5158 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5159 {
5160         unsigned long i;
5161         struct page *page;
5162         unsigned long num_pages;
5163
5164         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5165         num_pages = num_extent_pages(eb->start, eb->len);
5166         for (i = 0; i < num_pages; i++) {
5167                 page = eb->pages[i];
5168                 SetPageUptodate(page);
5169         }
5170 }
5171
5172 int extent_buffer_uptodate(struct extent_buffer *eb)
5173 {
5174         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5175 }
5176
5177 int read_extent_buffer_pages(struct extent_io_tree *tree,
5178                              struct extent_buffer *eb, u64 start, int wait,
5179                              get_extent_t *get_extent, int mirror_num)
5180 {
5181         unsigned long i;
5182         unsigned long start_i;
5183         struct page *page;
5184         int err;
5185         int ret = 0;
5186         int locked_pages = 0;
5187         int all_uptodate = 1;
5188         unsigned long num_pages;
5189         unsigned long num_reads = 0;
5190         struct bio *bio = NULL;
5191         unsigned long bio_flags = 0;
5192
5193         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5194                 return 0;
5195
5196         if (start) {
5197                 WARN_ON(start < eb->start);
5198                 start_i = (start >> PAGE_CACHE_SHIFT) -
5199                         (eb->start >> PAGE_CACHE_SHIFT);
5200         } else {
5201                 start_i = 0;
5202         }
5203
5204         num_pages = num_extent_pages(eb->start, eb->len);
5205         for (i = start_i; i < num_pages; i++) {
5206                 page = eb->pages[i];
5207                 if (wait == WAIT_NONE) {
5208                         if (!trylock_page(page))
5209                                 goto unlock_exit;
5210                 } else {
5211                         lock_page(page);
5212                 }
5213                 locked_pages++;
5214                 if (!PageUptodate(page)) {
5215                         num_reads++;
5216                         all_uptodate = 0;
5217                 }
5218         }
5219         if (all_uptodate) {
5220                 if (start_i == 0)
5221                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5222                 goto unlock_exit;
5223         }
5224
5225         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5226         eb->read_mirror = 0;
5227         atomic_set(&eb->io_pages, num_reads);
5228         for (i = start_i; i < num_pages; i++) {
5229                 page = eb->pages[i];
5230                 if (!PageUptodate(page)) {
5231                         ClearPageError(page);
5232                         err = __extent_read_full_page(tree, page,
5233                                                       get_extent, &bio,
5234                                                       mirror_num, &bio_flags,
5235                                                       READ | REQ_META);
5236                         if (err)
5237                                 ret = err;
5238                 } else {
5239                         unlock_page(page);
5240                 }
5241         }
5242
5243         if (bio) {
5244                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
5245                                      bio_flags);
5246                 if (err)
5247                         return err;
5248         }
5249
5250         if (ret || wait != WAIT_COMPLETE)
5251                 return ret;
5252
5253         for (i = start_i; i < num_pages; i++) {
5254                 page = eb->pages[i];
5255                 wait_on_page_locked(page);
5256                 if (!PageUptodate(page))
5257                         ret = -EIO;
5258         }
5259
5260         return ret;
5261
5262 unlock_exit:
5263         i = start_i;
5264         while (locked_pages > 0) {
5265                 page = eb->pages[i];
5266                 i++;
5267                 unlock_page(page);
5268                 locked_pages--;
5269         }
5270         return ret;
5271 }
5272
5273 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5274                         unsigned long start,
5275                         unsigned long len)
5276 {
5277         size_t cur;
5278         size_t offset;
5279         struct page *page;
5280         char *kaddr;
5281         char *dst = (char *)dstv;
5282         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5283         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5284
5285         WARN_ON(start > eb->len);
5286         WARN_ON(start + len > eb->start + eb->len);
5287
5288         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5289
5290         while (len > 0) {
5291                 page = eb->pages[i];
5292
5293                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5294                 kaddr = page_address(page);
5295                 memcpy(dst, kaddr + offset, cur);
5296
5297                 dst += cur;
5298                 len -= cur;
5299                 offset = 0;
5300                 i++;
5301         }
5302 }
5303
5304 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5305                         unsigned long start,
5306                         unsigned long len)
5307 {
5308         size_t cur;
5309         size_t offset;
5310         struct page *page;
5311         char *kaddr;
5312         char __user *dst = (char __user *)dstv;
5313         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5314         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5315         int ret = 0;
5316
5317         WARN_ON(start > eb->len);
5318         WARN_ON(start + len > eb->start + eb->len);
5319
5320         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5321
5322         while (len > 0) {
5323                 page = eb->pages[i];
5324
5325                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5326                 kaddr = page_address(page);
5327                 if (copy_to_user(dst, kaddr + offset, cur)) {
5328                         ret = -EFAULT;
5329                         break;
5330                 }
5331
5332                 dst += cur;
5333                 len -= cur;
5334                 offset = 0;
5335                 i++;
5336         }
5337
5338         return ret;
5339 }
5340
5341 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5342                                unsigned long min_len, char **map,
5343                                unsigned long *map_start,
5344                                unsigned long *map_len)
5345 {
5346         size_t offset = start & (PAGE_CACHE_SIZE - 1);
5347         char *kaddr;
5348         struct page *p;
5349         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5350         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5351         unsigned long end_i = (start_offset + start + min_len - 1) >>
5352                 PAGE_CACHE_SHIFT;
5353
5354         if (i != end_i)
5355                 return -EINVAL;
5356
5357         if (i == 0) {
5358                 offset = start_offset;
5359                 *map_start = 0;
5360         } else {
5361                 offset = 0;
5362                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
5363         }
5364
5365         if (start + min_len > eb->len) {
5366                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
5367                        "wanted %lu %lu\n",
5368                        eb->start, eb->len, start, min_len);
5369                 return -EINVAL;
5370         }
5371
5372         p = eb->pages[i];
5373         kaddr = page_address(p);
5374         *map = kaddr + offset;
5375         *map_len = PAGE_CACHE_SIZE - offset;
5376         return 0;
5377 }
5378
5379 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5380                           unsigned long start,
5381                           unsigned long len)
5382 {
5383         size_t cur;
5384         size_t offset;
5385         struct page *page;
5386         char *kaddr;
5387         char *ptr = (char *)ptrv;
5388         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5389         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5390         int ret = 0;
5391
5392         WARN_ON(start > eb->len);
5393         WARN_ON(start + len > eb->start + eb->len);
5394
5395         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5396
5397         while (len > 0) {
5398                 page = eb->pages[i];
5399
5400                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5401
5402                 kaddr = page_address(page);
5403                 ret = memcmp(ptr, kaddr + offset, cur);
5404                 if (ret)
5405                         break;
5406
5407                 ptr += cur;
5408                 len -= cur;
5409                 offset = 0;
5410                 i++;
5411         }
5412         return ret;
5413 }
5414
5415 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5416                          unsigned long start, unsigned long len)
5417 {
5418         size_t cur;
5419         size_t offset;
5420         struct page *page;
5421         char *kaddr;
5422         char *src = (char *)srcv;
5423         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5424         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5425
5426         WARN_ON(start > eb->len);
5427         WARN_ON(start + len > eb->start + eb->len);
5428
5429         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5430
5431         while (len > 0) {
5432                 page = eb->pages[i];
5433                 WARN_ON(!PageUptodate(page));
5434
5435                 cur = min(len, PAGE_CACHE_SIZE - offset);
5436                 kaddr = page_address(page);
5437                 memcpy(kaddr + offset, src, cur);
5438
5439                 src += cur;
5440                 len -= cur;
5441                 offset = 0;
5442                 i++;
5443         }
5444 }
5445
5446 void memset_extent_buffer(struct extent_buffer *eb, char c,
5447                           unsigned long start, unsigned long len)
5448 {
5449         size_t cur;
5450         size_t offset;
5451         struct page *page;
5452         char *kaddr;
5453         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5454         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5455
5456         WARN_ON(start > eb->len);
5457         WARN_ON(start + len > eb->start + eb->len);
5458
5459         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5460
5461         while (len > 0) {
5462                 page = eb->pages[i];
5463                 WARN_ON(!PageUptodate(page));
5464
5465                 cur = min(len, PAGE_CACHE_SIZE - offset);
5466                 kaddr = page_address(page);
5467                 memset(kaddr + offset, c, cur);
5468
5469                 len -= cur;
5470                 offset = 0;
5471                 i++;
5472         }
5473 }
5474
5475 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5476                         unsigned long dst_offset, unsigned long src_offset,
5477                         unsigned long len)
5478 {
5479         u64 dst_len = dst->len;
5480         size_t cur;
5481         size_t offset;
5482         struct page *page;
5483         char *kaddr;
5484         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5485         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5486
5487         WARN_ON(src->len != dst_len);
5488
5489         offset = (start_offset + dst_offset) &
5490                 (PAGE_CACHE_SIZE - 1);
5491
5492         while (len > 0) {
5493                 page = dst->pages[i];
5494                 WARN_ON(!PageUptodate(page));
5495
5496                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5497
5498                 kaddr = page_address(page);
5499                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5500
5501                 src_offset += cur;
5502                 len -= cur;
5503                 offset = 0;
5504                 i++;
5505         }
5506 }
5507
5508 /*
5509  * The extent buffer bitmap operations are done with byte granularity because
5510  * bitmap items are not guaranteed to be aligned to a word and therefore a
5511  * single word in a bitmap may straddle two pages in the extent buffer.
5512  */
5513 #define BIT_BYTE(nr) ((nr) / BITS_PER_BYTE)
5514 #define BYTE_MASK ((1 << BITS_PER_BYTE) - 1)
5515 #define BITMAP_FIRST_BYTE_MASK(start) \
5516         ((BYTE_MASK << ((start) & (BITS_PER_BYTE - 1))) & BYTE_MASK)
5517 #define BITMAP_LAST_BYTE_MASK(nbits) \
5518         (BYTE_MASK >> (-(nbits) & (BITS_PER_BYTE - 1)))
5519
5520 /*
5521  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5522  * given bit number
5523  * @eb: the extent buffer
5524  * @start: offset of the bitmap item in the extent buffer
5525  * @nr: bit number
5526  * @page_index: return index of the page in the extent buffer that contains the
5527  * given bit number
5528  * @page_offset: return offset into the page given by page_index
5529  *
5530  * This helper hides the ugliness of finding the byte in an extent buffer which
5531  * contains a given bit.
5532  */
5533 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5534                                     unsigned long start, unsigned long nr,
5535                                     unsigned long *page_index,
5536                                     size_t *page_offset)
5537 {
5538         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5539         size_t byte_offset = BIT_BYTE(nr);
5540         size_t offset;
5541
5542         /*
5543          * The byte we want is the offset of the extent buffer + the offset of
5544          * the bitmap item in the extent buffer + the offset of the byte in the
5545          * bitmap item.
5546          */
5547         offset = start_offset + start + byte_offset;
5548
5549         *page_index = offset >> PAGE_CACHE_SHIFT;
5550         *page_offset = offset & (PAGE_CACHE_SIZE - 1);
5551 }
5552
5553 /**
5554  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5555  * @eb: the extent buffer
5556  * @start: offset of the bitmap item in the extent buffer
5557  * @nr: bit number to test
5558  */
5559 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5560                            unsigned long nr)
5561 {
5562         char *kaddr;
5563         struct page *page;
5564         unsigned long i;
5565         size_t offset;
5566
5567         eb_bitmap_offset(eb, start, nr, &i, &offset);
5568         page = eb->pages[i];
5569         WARN_ON(!PageUptodate(page));
5570         kaddr = page_address(page);
5571         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5572 }
5573
5574 /**
5575  * extent_buffer_bitmap_set - set an area of a bitmap
5576  * @eb: the extent buffer
5577  * @start: offset of the bitmap item in the extent buffer
5578  * @pos: bit number of the first bit
5579  * @len: number of bits to set
5580  */
5581 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5582                               unsigned long pos, unsigned long len)
5583 {
5584         char *kaddr;
5585         struct page *page;
5586         unsigned long i;
5587         size_t offset;
5588         const unsigned int size = pos + len;
5589         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5590         unsigned int mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5591
5592         eb_bitmap_offset(eb, start, pos, &i, &offset);
5593         page = eb->pages[i];
5594         WARN_ON(!PageUptodate(page));
5595         kaddr = page_address(page);
5596
5597         while (len >= bits_to_set) {
5598                 kaddr[offset] |= mask_to_set;
5599                 len -= bits_to_set;
5600                 bits_to_set = BITS_PER_BYTE;
5601                 mask_to_set = ~0U;
5602                 if (++offset >= PAGE_CACHE_SIZE && len > 0) {
5603                         offset = 0;
5604                         page = eb->pages[++i];
5605                         WARN_ON(!PageUptodate(page));
5606                         kaddr = page_address(page);
5607                 }
5608         }
5609         if (len) {
5610                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5611                 kaddr[offset] |= mask_to_set;
5612         }
5613 }
5614
5615
5616 /**
5617  * extent_buffer_bitmap_clear - clear an area of a bitmap
5618  * @eb: the extent buffer
5619  * @start: offset of the bitmap item in the extent buffer
5620  * @pos: bit number of the first bit
5621  * @len: number of bits to clear
5622  */
5623 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5624                                 unsigned long pos, unsigned long len)
5625 {
5626         char *kaddr;
5627         struct page *page;
5628         unsigned long i;
5629         size_t offset;
5630         const unsigned int size = pos + len;
5631         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5632         unsigned int mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5633
5634         eb_bitmap_offset(eb, start, pos, &i, &offset);
5635         page = eb->pages[i];
5636         WARN_ON(!PageUptodate(page));
5637         kaddr = page_address(page);
5638
5639         while (len >= bits_to_clear) {
5640                 kaddr[offset] &= ~mask_to_clear;
5641                 len -= bits_to_clear;
5642                 bits_to_clear = BITS_PER_BYTE;
5643                 mask_to_clear = ~0U;
5644                 if (++offset >= PAGE_CACHE_SIZE && len > 0) {
5645                         offset = 0;
5646                         page = eb->pages[++i];
5647                         WARN_ON(!PageUptodate(page));
5648                         kaddr = page_address(page);
5649                 }
5650         }
5651         if (len) {
5652                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5653                 kaddr[offset] &= ~mask_to_clear;
5654         }
5655 }
5656
5657 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5658 {
5659         unsigned long distance = (src > dst) ? src - dst : dst - src;
5660         return distance < len;
5661 }
5662
5663 static void copy_pages(struct page *dst_page, struct page *src_page,
5664                        unsigned long dst_off, unsigned long src_off,
5665                        unsigned long len)
5666 {
5667         char *dst_kaddr = page_address(dst_page);
5668         char *src_kaddr;
5669         int must_memmove = 0;
5670
5671         if (dst_page != src_page) {
5672                 src_kaddr = page_address(src_page);
5673         } else {
5674                 src_kaddr = dst_kaddr;
5675                 if (areas_overlap(src_off, dst_off, len))
5676                         must_memmove = 1;
5677         }
5678
5679         if (must_memmove)
5680                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5681         else
5682                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5683 }
5684
5685 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5686                            unsigned long src_offset, unsigned long len)
5687 {
5688         size_t cur;
5689         size_t dst_off_in_page;
5690         size_t src_off_in_page;
5691         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5692         unsigned long dst_i;
5693         unsigned long src_i;
5694
5695         if (src_offset + len > dst->len) {
5696                 btrfs_err(dst->fs_info,
5697                         "memmove bogus src_offset %lu move "
5698                        "len %lu dst len %lu", src_offset, len, dst->len);
5699                 BUG_ON(1);
5700         }
5701         if (dst_offset + len > dst->len) {
5702                 btrfs_err(dst->fs_info,
5703                         "memmove bogus dst_offset %lu move "
5704                        "len %lu dst len %lu", dst_offset, len, dst->len);
5705                 BUG_ON(1);
5706         }
5707
5708         while (len > 0) {
5709                 dst_off_in_page = (start_offset + dst_offset) &
5710                         (PAGE_CACHE_SIZE - 1);
5711                 src_off_in_page = (start_offset + src_offset) &
5712                         (PAGE_CACHE_SIZE - 1);
5713
5714                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5715                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5716
5717                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5718                                                src_off_in_page));
5719                 cur = min_t(unsigned long, cur,
5720                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5721
5722                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5723                            dst_off_in_page, src_off_in_page, cur);
5724
5725                 src_offset += cur;
5726                 dst_offset += cur;
5727                 len -= cur;
5728         }
5729 }
5730
5731 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5732                            unsigned long src_offset, unsigned long len)
5733 {
5734         size_t cur;
5735         size_t dst_off_in_page;
5736         size_t src_off_in_page;
5737         unsigned long dst_end = dst_offset + len - 1;
5738         unsigned long src_end = src_offset + len - 1;
5739         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5740         unsigned long dst_i;
5741         unsigned long src_i;
5742
5743         if (src_offset + len > dst->len) {
5744                 btrfs_err(dst->fs_info, "memmove bogus src_offset %lu move "
5745                        "len %lu len %lu", src_offset, len, dst->len);
5746                 BUG_ON(1);
5747         }
5748         if (dst_offset + len > dst->len) {
5749                 btrfs_err(dst->fs_info, "memmove bogus dst_offset %lu move "
5750                        "len %lu len %lu", dst_offset, len, dst->len);
5751                 BUG_ON(1);
5752         }
5753         if (dst_offset < src_offset) {
5754                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5755                 return;
5756         }
5757         while (len > 0) {
5758                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5759                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5760
5761                 dst_off_in_page = (start_offset + dst_end) &
5762                         (PAGE_CACHE_SIZE - 1);
5763                 src_off_in_page = (start_offset + src_end) &
5764                         (PAGE_CACHE_SIZE - 1);
5765
5766                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5767                 cur = min(cur, dst_off_in_page + 1);
5768                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5769                            dst_off_in_page - cur + 1,
5770                            src_off_in_page - cur + 1, cur);
5771
5772                 dst_end -= cur;
5773                 src_end -= cur;
5774                 len -= cur;
5775         }
5776 }
5777
5778 int try_release_extent_buffer(struct page *page)
5779 {
5780         struct extent_buffer *eb;
5781
5782         /*
5783          * We need to make sure noboody is attaching this page to an eb right
5784          * now.
5785          */
5786         spin_lock(&page->mapping->private_lock);
5787         if (!PagePrivate(page)) {
5788                 spin_unlock(&page->mapping->private_lock);
5789                 return 1;
5790         }
5791
5792         eb = (struct extent_buffer *)page->private;
5793         BUG_ON(!eb);
5794
5795         /*
5796          * This is a little awful but should be ok, we need to make sure that
5797          * the eb doesn't disappear out from under us while we're looking at
5798          * this page.
5799          */
5800         spin_lock(&eb->refs_lock);
5801         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5802                 spin_unlock(&eb->refs_lock);
5803                 spin_unlock(&page->mapping->private_lock);
5804                 return 0;
5805         }
5806         spin_unlock(&page->mapping->private_lock);
5807
5808         /*
5809          * If tree ref isn't set then we know the ref on this eb is a real ref,
5810          * so just return, this page will likely be freed soon anyway.
5811          */
5812         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5813                 spin_unlock(&eb->refs_lock);
5814                 return 0;
5815         }
5816
5817         return release_extent_buffer(eb);
5818 }