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         kmem_cache_destroy(extent_state_cache);
 210         kmem_cache_destroy(extent_buffer_cache);
 211         if (btrfs_bioset)
 212                 bioset_free(btrfs_bioset);
 213 }
 214
 215 void extent_io_tree_init(struct extent_io_tree *tree,
 216                          struct address_space *mapping)
 217 {
 218         tree->state = RB_ROOT;
 219         tree->ops = NULL;
 220         tree->dirty_bytes = 0;
 221         spin_lock_init(&tree->lock);
 222         tree->mapping = mapping;
 223 }
 224
 225 static struct extent_state *alloc_extent_state(gfp_t mask)
 226 {
 227         struct extent_state *state;
 228
 229         state = kmem_cache_alloc(extent_state_cache, mask);
 230         if (!state)
 231                 return state;
 232         state->state = 0;
 233         state->failrec = NULL;
 234         RB_CLEAR_NODE(&state->rb_node);
 235         btrfs_leak_debug_add(&state->leak_list, &states);
 236         atomic_set(&state->refs, 1);
 237         init_waitqueue_head(&state->wq);
 238         trace_alloc_extent_state(state, mask, _RET_IP_);
 239         return state;
 240 }
 241
 242 void free_extent_state(struct extent_state *state)
 243 {
 244         if (!state)
 245                 return;
 246         if (atomic_dec_and_test(&state->refs)) {
 247                 WARN_ON(extent_state_in_tree(state));
 248                 btrfs_leak_debug_del(&state->leak_list);
 249                 trace_free_extent_state(state, _RET_IP_);
 250                 kmem_cache_free(extent_state_cache, state);
 251         }
 252 }
 253
 254 static struct rb_node *tree_insert(struct rb_root *root,
 255                                    struct rb_node *search_start,
 256                                    u64 offset,
 257                                    struct rb_node *node,
 258                                    struct rb_node ***p_in,
 259                                    struct rb_node **parent_in)
 260 {
 261         struct rb_node **p;
 262         struct rb_node *parent = NULL;
 263         struct tree_entry *entry;
 264
 265         if (p_in && parent_in) {
 266                 p = *p_in;
 267                 parent = *parent_in;
 268                 goto do_insert;
 269         }
 270
 271         p = search_start ? &search_start : &root->rb_node;
 272         while (*p) {
 273                 parent = *p;
 274                 entry = rb_entry(parent, struct tree_entry, rb_node);
 275
 276                 if (offset < entry->start)
 277                         p = &(*p)->rb_left;
 278                 else if (offset > entry->end)
 279                         p = &(*p)->rb_right;
 280                 else
 281                         return parent;
 282         }
 283
 284 do_insert:
 285         rb_link_node(node, parent, p);
 286         rb_insert_color(node, root);
 287         return NULL;
 288 }
 289
 290 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 291                                       struct rb_node **prev_ret,
 292                                       struct rb_node **next_ret,
 293                                       struct rb_node ***p_ret,
 294                                       struct rb_node **parent_ret)
 295 {
 296         struct rb_root *root = &tree->state;
 297         struct rb_node **n = &root->rb_node;
 298         struct rb_node *prev = NULL;
 299         struct rb_node *orig_prev = NULL;
 300         struct tree_entry *entry;
 301         struct tree_entry *prev_entry = NULL;
 302
 303         while (*n) {
 304                 prev = *n;
 305                 entry = rb_entry(prev, struct tree_entry, rb_node);
 306                 prev_entry = entry;
 307
 308                 if (offset < entry->start)
 309                         n = &(*n)->rb_left;
 310                 else if (offset > entry->end)
 311                         n = &(*n)->rb_right;
 312                 else
 313                         return *n;
 314         }
 315
 316         if (p_ret)
 317                 *p_ret = n;
 318         if (parent_ret)
 319                 *parent_ret = prev;
 320
 321         if (prev_ret) {
 322                 orig_prev = prev;
 323                 while (prev && offset > prev_entry->end) {
 324                         prev = rb_next(prev);
 325                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 326                 }
 327                 *prev_ret = prev;
 328                 prev = orig_prev;
 329         }
 330
 331         if (next_ret) {
 332                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 333                 while (prev && offset < prev_entry->start) {
 334                         prev = rb_prev(prev);
 335                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 336                 }
 337                 *next_ret = prev;
 338         }
 339         return NULL;
 340 }
 341
 342 static inline struct rb_node *
 343 tree_search_for_insert(struct extent_io_tree *tree,
 344                        u64 offset,
 345                        struct rb_node ***p_ret,
 346                        struct rb_node **parent_ret)
 347 {
 348         struct rb_node *prev = NULL;
 349         struct rb_node *ret;
 350
 351         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 352         if (!ret)
 353                 return prev;
 354         return ret;
 355 }
 356
 357 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 358                                           u64 offset)
 359 {
 360         return tree_search_for_insert(tree, offset, NULL, NULL);
 361 }
 362
 363 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 364                      struct extent_state *other)
 365 {
 366         if (tree->ops && tree->ops->merge_extent_hook)
 367                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 368                                              other);
 369 }
 370
 371 /*
 372  * utility function to look for merge candidates inside a given range.
 373  * Any extents with matching state are merged together into a single
 374  * extent in the tree.  Extents with EXTENT_IO in their state field
 375  * are not merged because the end_io handlers need to be able to do
 376  * operations on them without sleeping (or doing allocations/splits).
 377  *
 378  * This should be called with the tree lock held.
 379  */
 380 static void merge_state(struct extent_io_tree *tree,
 381                         struct extent_state *state)
 382 {
 383         struct extent_state *other;
 384         struct rb_node *other_node;
 385
 386         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 387                 return;
 388
 389         other_node = rb_prev(&state->rb_node);
 390         if (other_node) {
 391                 other = rb_entry(other_node, struct extent_state, rb_node);
 392                 if (other->end == state->start - 1 &&
 393                     other->state == state->state) {
 394                         merge_cb(tree, state, other);
 395                         state->start = other->start;
 396                         rb_erase(&other->rb_node, &tree->state);
 397                         RB_CLEAR_NODE(&other->rb_node);
 398                         free_extent_state(other);
 399                 }
 400         }
 401         other_node = rb_next(&state->rb_node);
 402         if (other_node) {
 403                 other = rb_entry(other_node, struct extent_state, rb_node);
 404                 if (other->start == state->end + 1 &&
 405                     other->state == state->state) {
 406                         merge_cb(tree, state, other);
 407                         state->end = other->end;
 408                         rb_erase(&other->rb_node, &tree->state);
 409                         RB_CLEAR_NODE(&other->rb_node);
 410                         free_extent_state(other);
 411                 }
 412         }
 413 }
 414
 415 static void set_state_cb(struct extent_io_tree *tree,
 416                          struct extent_state *state, unsigned *bits)
 417 {
 418         if (tree->ops && tree->ops->set_bit_hook)
 419                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 420 }
 421
 422 static void clear_state_cb(struct extent_io_tree *tree,
 423                            struct extent_state *state, unsigned *bits)
 424 {
 425         if (tree->ops && tree->ops->clear_bit_hook)
 426                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 427 }
 428
 429 static void set_state_bits(struct extent_io_tree *tree,
 430                            struct extent_state *state, unsigned *bits,
 431                            struct extent_changeset *changeset);
 432
 433 /*
 434  * insert an extent_state struct into the tree.  'bits' are set on the
 435  * struct before it is inserted.
 436  *
 437  * This may return -EEXIST if the extent is already there, in which case the
 438  * state struct is freed.
 439  *
 440  * The tree lock is not taken internally.  This is a utility function and
 441  * probably isn't what you want to call (see set/clear_extent_bit).
 442  */
 443 static int insert_state(struct extent_io_tree *tree,
 444                         struct extent_state *state, u64 start, u64 end,
 445                         struct rb_node ***p,
 446                         struct rb_node **parent,
 447                         unsigned *bits, struct extent_changeset *changeset)
 448 {
 449         struct rb_node *node;
 450
 451         if (end < start)
 452                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 453                        end, start);
 454         state->start = start;
 455         state->end = end;
 456
 457         set_state_bits(tree, state, bits, changeset);
 458
 459         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 460         if (node) {
 461                 struct extent_state *found;
 462                 found = rb_entry(node, struct extent_state, rb_node);
 463                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
 464                        "%llu %llu\n",
 465                        found->start, found->end, start, end);
 466                 return -EEXIST;
 467         }
 468         merge_state(tree, state);
 469         return 0;
 470 }
 471
 472 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 473                      u64 split)
 474 {
 475         if (tree->ops && tree->ops->split_extent_hook)
 476                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 477 }
 478
 479 /*
 480  * split a given extent state struct in two, inserting the preallocated
 481  * struct 'prealloc' as the newly created second half.  'split' indicates an
 482  * offset inside 'orig' where it should be split.
 483  *
 484  * Before calling,
 485  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 486  * are two extent state structs in the tree:
 487  * prealloc: [orig->start, split - 1]
 488  * orig: [ split, orig->end ]
 489  *
 490  * The tree locks are not taken by this function. They need to be held
 491  * by the caller.
 492  */
 493 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 494                        struct extent_state *prealloc, u64 split)
 495 {
 496         struct rb_node *node;
 497
 498         split_cb(tree, orig, split);
 499
 500         prealloc->start = orig->start;
 501         prealloc->end = split - 1;
 502         prealloc->state = orig->state;
 503         orig->start = split;
 504
 505         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 506                            &prealloc->rb_node, NULL, NULL);
 507         if (node) {
 508                 free_extent_state(prealloc);
 509                 return -EEXIST;
 510         }
 511         return 0;
 512 }
 513
 514 static struct extent_state *next_state(struct extent_state *state)
 515 {
 516         struct rb_node *next = rb_next(&state->rb_node);
 517         if (next)
 518                 return rb_entry(next, struct extent_state, rb_node);
 519         else
 520                 return NULL;
 521 }
 522
 523 /*
 524  * utility function to clear some bits in an extent state struct.
 525  * it will optionally wake up any one waiting on this state (wake == 1).
 526  *
 527  * If no bits are set on the state struct after clearing things, the
 528  * struct is freed and removed from the tree
 529  */
 530 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 531                                             struct extent_state *state,
 532                                             unsigned *bits, int wake,
 533                                             struct extent_changeset *changeset)
 534 {
 535         struct extent_state *next;
 536         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 537
 538         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 539                 u64 range = state->end - state->start + 1;
 540                 WARN_ON(range > tree->dirty_bytes);
 541                 tree->dirty_bytes -= range;
 542         }
 543         clear_state_cb(tree, state, bits);
 544         add_extent_changeset(state, bits_to_clear, changeset, 0);
 545         state->state &= ~bits_to_clear;
 546         if (wake)
 547                 wake_up(&state->wq);
 548         if (state->state == 0) {
 549                 next = next_state(state);
 550                 if (extent_state_in_tree(state)) {
 551                         rb_erase(&state->rb_node, &tree->state);
 552                         RB_CLEAR_NODE(&state->rb_node);
 553                         free_extent_state(state);
 554                 } else {
 555                         WARN_ON(1);
 556                 }
 557         } else {
 558                 merge_state(tree, state);
 559                 next = next_state(state);
 560         }
 561         return next;
 562 }
 563
 564 static struct extent_state *
 565 alloc_extent_state_atomic(struct extent_state *prealloc)
 566 {
 567         if (!prealloc)
 568                 prealloc = alloc_extent_state(GFP_ATOMIC);
 569
 570         return prealloc;
 571 }
 572
 573 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 574 {
 575         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 576                     "Extent tree was modified by another "
 577                     "thread while locked.");
 578 }
 579
 580 /*
 581  * clear some bits on a range in the tree.  This may require splitting
 582  * or inserting elements in the tree, so the gfp mask is used to
 583  * indicate which allocations or sleeping are allowed.
 584  *
 585  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 586  * the given range from the tree regardless of state (ie for truncate).
 587  *
 588  * the range [start, end] is inclusive.
 589  *
 590  * This takes the tree lock, and returns 0 on success and < 0 on error.
 591  */
 592 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 593                               unsigned bits, int wake, int delete,
 594                               struct extent_state **cached_state,
 595                               gfp_t mask, struct extent_changeset *changeset)
 596 {
 597         struct extent_state *state;
 598         struct extent_state *cached;
 599         struct extent_state *prealloc = NULL;
 600         struct rb_node *node;
 601         u64 last_end;
 602         int err;
 603         int clear = 0;
 604
 605         btrfs_debug_check_extent_io_range(tree, start, end);
 606
 607         if (bits & EXTENT_DELALLOC)
 608                 bits |= EXTENT_NORESERVE;
 609
 610         if (delete)
 611                 bits |= ~EXTENT_CTLBITS;
 612         bits |= EXTENT_FIRST_DELALLOC;
 613
 614         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 615                 clear = 1;
 616 again:
 617         if (!prealloc && gfpflags_allow_blocking(mask)) {
 618                 /*
 619                  * Don't care for allocation failure here because we might end
 620                  * up not needing the pre-allocated extent state at all, which
 621                  * is the case if we only have in the tree extent states that
 622                  * cover our input range and don't cover too any other range.
 623                  * If we end up needing a new extent state we allocate it later.
 624                  */
 625                 prealloc = alloc_extent_state(mask);
 626         }
 627
 628         spin_lock(&tree->lock);
 629         if (cached_state) {
 630                 cached = *cached_state;
 631
 632                 if (clear) {
 633                         *cached_state = NULL;
 634                         cached_state = NULL;
 635                 }
 636
 637                 if (cached && extent_state_in_tree(cached) &&
 638                     cached->start <= start && cached->end > start) {
 639                         if (clear)
 640                                 atomic_dec(&cached->refs);
 641                         state = cached;
 642                         goto hit_next;
 643                 }
 644                 if (clear)
 645                         free_extent_state(cached);
 646         }
 647         /*
 648          * this search will find the extents that end after
 649          * our range starts
 650          */
 651         node = tree_search(tree, start);
 652         if (!node)
 653                 goto out;
 654         state = rb_entry(node, struct extent_state, rb_node);
 655 hit_next:
 656         if (state->start > end)
 657                 goto out;
 658         WARN_ON(state->end < start);
 659         last_end = state->end;
 660
 661         /* the state doesn't have the wanted bits, go ahead */
 662         if (!(state->state & bits)) {
 663                 state = next_state(state);
 664                 goto next;
 665         }
 666
 667         /*
 668          *     | ---- desired range ---- |
 669          *  | state | or
 670          *  | ------------- state -------------- |
 671          *
 672          * We need to split the extent we found, and may flip
 673          * bits on second half.
 674          *
 675          * If the extent we found extends past our range, we
 676          * just split and search again.  It'll get split again
 677          * the next time though.
 678          *
 679          * If the extent we found is inside our range, we clear
 680          * the desired bit on it.
 681          */
 682
 683         if (state->start < start) {
 684                 prealloc = alloc_extent_state_atomic(prealloc);
 685                 BUG_ON(!prealloc);
 686                 err = split_state(tree, state, prealloc, start);
 687                 if (err)
 688                         extent_io_tree_panic(tree, err);
 689
 690                 prealloc = NULL;
 691                 if (err)
 692                         goto out;
 693                 if (state->end <= end) {
 694                         state = clear_state_bit(tree, state, &bits, wake,
 695                                                 changeset);
 696                         goto next;
 697                 }
 698                 goto search_again;
 699         }
 700         /*
 701          * | ---- desired range ---- |
 702          *                        | state |
 703          * We need to split the extent, and clear the bit
 704          * on the first half
 705          */
 706         if (state->start <= end && state->end > end) {
 707                 prealloc = alloc_extent_state_atomic(prealloc);
 708                 BUG_ON(!prealloc);
 709                 err = split_state(tree, state, prealloc, end + 1);
 710                 if (err)
 711                         extent_io_tree_panic(tree, err);
 712
 713                 if (wake)
 714                         wake_up(&state->wq);
 715
 716                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 717
 718                 prealloc = NULL;
 719                 goto out;
 720         }
 721
 722         state = clear_state_bit(tree, state, &bits, wake, changeset);
 723 next:
 724         if (last_end == (u64)-1)
 725                 goto out;
 726         start = last_end + 1;
 727         if (start <= end && state && !need_resched())
 728                 goto hit_next;
 729         goto search_again;
 730
 731 out:
 732         spin_unlock(&tree->lock);
 733         if (prealloc)
 734                 free_extent_state(prealloc);
 735
 736         return 0;
 737
 738 search_again:
 739         if (start > end)
 740                 goto out;
 741         spin_unlock(&tree->lock);
 742         if (gfpflags_allow_blocking(mask))
 743                 cond_resched();
 744         goto again;
 745 }
 746
 747 static void wait_on_state(struct extent_io_tree *tree,
 748                           struct extent_state *state)
 749                 __releases(tree->lock)
 750                 __acquires(tree->lock)
 751 {
 752         DEFINE_WAIT(wait);
 753         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 754         spin_unlock(&tree->lock);
 755         schedule();
 756         spin_lock(&tree->lock);
 757         finish_wait(&state->wq, &wait);
 758 }
 759
 760 /*
 761  * waits for one or more bits to clear on a range in the state tree.
 762  * The range [start, end] is inclusive.
 763  * The tree lock is taken by this function
 764  */
 765 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 766                             unsigned long bits)
 767 {
 768         struct extent_state *state;
 769         struct rb_node *node;
 770
 771         btrfs_debug_check_extent_io_range(tree, start, end);
 772
 773         spin_lock(&tree->lock);
 774 again:
 775         while (1) {
 776                 /*
 777                  * this search will find all the extents that end after
 778                  * our range starts
 779                  */
 780                 node = tree_search(tree, start);
 781 process_node:
 782                 if (!node)
 783                         break;
 784
 785                 state = rb_entry(node, struct extent_state, rb_node);
 786
 787                 if (state->start > end)
 788                         goto out;
 789
 790                 if (state->state & bits) {
 791                         start = state->start;
 792                         atomic_inc(&state->refs);
 793                         wait_on_state(tree, state);
 794                         free_extent_state(state);
 795                         goto again;
 796                 }
 797                 start = state->end + 1;
 798
 799                 if (start > end)
 800                         break;
 801
 802                 if (!cond_resched_lock(&tree->lock)) {
 803                         node = rb_next(node);
 804                         goto process_node;
 805                 }
 806         }
 807 out:
 808         spin_unlock(&tree->lock);
 809 }
 810
 811 static void set_state_bits(struct extent_io_tree *tree,
 812                            struct extent_state *state,
 813                            unsigned *bits, struct extent_changeset *changeset)
 814 {
 815         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 816
 817         set_state_cb(tree, state, bits);
 818         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 819                 u64 range = state->end - state->start + 1;
 820                 tree->dirty_bytes += range;
 821         }
 822         add_extent_changeset(state, bits_to_set, changeset, 1);
 823         state->state |= bits_to_set;
 824 }
 825
 826 static void cache_state_if_flags(struct extent_state *state,
 827                                  struct extent_state **cached_ptr,
 828                                  unsigned flags)
 829 {
 830         if (cached_ptr && !(*cached_ptr)) {
 831                 if (!flags || (state->state & flags)) {
 832                         *cached_ptr = state;
 833                         atomic_inc(&state->refs);
 834                 }
 835         }
 836 }
 837
 838 static void cache_state(struct extent_state *state,
 839                         struct extent_state **cached_ptr)
 840 {
 841         return cache_state_if_flags(state, cached_ptr,
 842                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 843 }
 844
 845 /*
 846  * set some bits on a range in the tree.  This may require allocations or
 847  * sleeping, so the gfp mask is used to indicate what is allowed.
 848  *
 849  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 850  * part of the range already has the desired bits set.  The start of the
 851  * existing range is returned in failed_start in this case.
 852  *
 853  * [start, end] is inclusive This takes the tree lock.
 854  */
 855
 856 static int __must_check
 857 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 858                  unsigned bits, unsigned exclusive_bits,
 859                  u64 *failed_start, struct extent_state **cached_state,
 860                  gfp_t mask, struct extent_changeset *changeset)
 861 {
 862         struct extent_state *state;
 863         struct extent_state *prealloc = NULL;
 864         struct rb_node *node;
 865         struct rb_node **p;
 866         struct rb_node *parent;
 867         int err = 0;
 868         u64 last_start;
 869         u64 last_end;
 870
 871         btrfs_debug_check_extent_io_range(tree, start, end);
 872
 873         bits |= EXTENT_FIRST_DELALLOC;
 874 again:
 875         if (!prealloc && gfpflags_allow_blocking(mask)) {
 876                 prealloc = alloc_extent_state(mask);
 877                 BUG_ON(!prealloc);
 878         }
 879
 880         spin_lock(&tree->lock);
 881         if (cached_state && *cached_state) {
 882                 state = *cached_state;
 883                 if (state->start <= start && state->end > start &&
 884                     extent_state_in_tree(state)) {
 885                         node = &state->rb_node;
 886                         goto hit_next;
 887                 }
 888         }
 889         /*
 890          * this search will find all the extents that end after
 891          * our range starts.
 892          */
 893         node = tree_search_for_insert(tree, start, &p, &parent);
 894         if (!node) {
 895                 prealloc = alloc_extent_state_atomic(prealloc);
 896                 BUG_ON(!prealloc);
 897                 err = insert_state(tree, prealloc, start, end,
 898                                    &p, &parent, &bits, changeset);
 899                 if (err)
 900                         extent_io_tree_panic(tree, err);
 901
 902                 cache_state(prealloc, cached_state);
 903                 prealloc = NULL;
 904                 goto out;
 905         }
 906         state = rb_entry(node, struct extent_state, rb_node);
 907 hit_next:
 908         last_start = state->start;
 909         last_end = state->end;
 910
 911         /*
 912          * | ---- desired range ---- |
 913          * | state |
 914          *
 915          * Just lock what we found and keep going
 916          */
 917         if (state->start == start && state->end <= end) {
 918                 if (state->state & exclusive_bits) {
 919                         *failed_start = state->start;
 920                         err = -EEXIST;
 921                         goto out;
 922                 }
 923
 924                 set_state_bits(tree, state, &bits, changeset);
 925                 cache_state(state, cached_state);
 926                 merge_state(tree, state);
 927                 if (last_end == (u64)-1)
 928                         goto out;
 929                 start = last_end + 1;
 930                 state = next_state(state);
 931                 if (start < end && state && state->start == start &&
 932                     !need_resched())
 933                         goto hit_next;
 934                 goto search_again;
 935         }
 936
 937         /*
 938          *     | ---- desired range ---- |
 939          * | state |
 940          *   or
 941          * | ------------- state -------------- |
 942          *
 943          * We need to split the extent we found, and may flip bits on
 944          * second half.
 945          *
 946          * If the extent we found extends past our
 947          * range, we just split and search again.  It'll get split
 948          * again the next time though.
 949          *
 950          * If the extent we found is inside our range, we set the
 951          * desired bit on it.
 952          */
 953         if (state->start < start) {
 954                 if (state->state & exclusive_bits) {
 955                         *failed_start = start;
 956                         err = -EEXIST;
 957                         goto out;
 958                 }
 959
 960                 prealloc = alloc_extent_state_atomic(prealloc);
 961                 BUG_ON(!prealloc);
 962                 err = split_state(tree, state, prealloc, start);
 963                 if (err)
 964                         extent_io_tree_panic(tree, err);
 965
 966                 prealloc = NULL;
 967                 if (err)
 968                         goto out;
 969                 if (state->end <= end) {
 970                         set_state_bits(tree, state, &bits, changeset);
 971                         cache_state(state, cached_state);
 972                         merge_state(tree, state);
 973                         if (last_end == (u64)-1)
 974                                 goto out;
 975                         start = last_end + 1;
 976                         state = next_state(state);
 977                         if (start < end && state && state->start == start &&
 978                             !need_resched())
 979                                 goto hit_next;
 980                 }
 981                 goto search_again;
 982         }
 983         /*
 984          * | ---- desired range ---- |
 985          *     | state | or               | state |
 986          *
 987          * There's a hole, we need to insert something in it and
 988          * ignore the extent we found.
 989          */
 990         if (state->start > start) {
 991                 u64 this_end;
 992                 if (end < last_start)
 993                         this_end = end;
 994                 else
 995                         this_end = last_start - 1;
 996
 997                 prealloc = alloc_extent_state_atomic(prealloc);
 998                 BUG_ON(!prealloc);
 999
1000                 /*
1001                  * Avoid to free 'prealloc' if it can be merged with
1002                  * the later extent.
1003                  */
1004                 err = insert_state(tree, prealloc, start, this_end,
1005                                    NULL, NULL, &bits, changeset);
1006                 if (err)
1007                         extent_io_tree_panic(tree, err);
1008
1009                 cache_state(prealloc, cached_state);
1010                 prealloc = NULL;
1011                 start = this_end + 1;
1012                 goto search_again;
1013         }
1014         /*
1015          * | ---- desired range ---- |
1016          *                        | state |
1017          * We need to split the extent, and set the bit
1018          * on the first half
1019          */
1020         if (state->start <= end && state->end > end) {
1021                 if (state->state & exclusive_bits) {
1022                         *failed_start = start;
1023                         err = -EEXIST;
1024                         goto out;
1025                 }
1026
1027                 prealloc = alloc_extent_state_atomic(prealloc);
1028                 BUG_ON(!prealloc);
1029                 err = split_state(tree, state, prealloc, end + 1);
1030                 if (err)
1031                         extent_io_tree_panic(tree, err);
1032
1033                 set_state_bits(tree, prealloc, &bits, changeset);
1034                 cache_state(prealloc, cached_state);
1035                 merge_state(tree, prealloc);
1036                 prealloc = NULL;
1037                 goto out;
1038         }
1039
1040         goto search_again;
1041
1042 out:
1043         spin_unlock(&tree->lock);
1044         if (prealloc)
1045                 free_extent_state(prealloc);
1046
1047         return err;
1048
1049 search_again:
1050         if (start > end)
1051                 goto out;
1052         spin_unlock(&tree->lock);
1053         if (gfpflags_allow_blocking(mask))
1054                 cond_resched();
1055         goto again;
1056 }
1057
1058 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1059                    unsigned bits, u64 * failed_start,
1060                    struct extent_state **cached_state, gfp_t mask)
1061 {
1062         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1063                                 cached_state, mask, NULL);
1064 }
1065
1066
1067 /**
1068  * convert_extent_bit - convert all bits in a given range from one bit to
1069  *                      another
1070  * @tree:       the io tree to search
1071  * @start:      the start offset in bytes
1072  * @end:        the end offset in bytes (inclusive)
1073  * @bits:       the bits to set in this range
1074  * @clear_bits: the bits to clear in this range
1075  * @cached_state:       state that we're going to cache
1076  * @mask:       the allocation mask
1077  *
1078  * This will go through and set bits for the given range.  If any states exist
1079  * already in this range they are set with the given bit and cleared of the
1080  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1081  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1082  * boundary bits like LOCK.
1083  */
1084 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1085                        unsigned bits, unsigned clear_bits,
1086                        struct extent_state **cached_state, gfp_t mask)
1087 {
1088         struct extent_state *state;
1089         struct extent_state *prealloc = NULL;
1090         struct rb_node *node;
1091         struct rb_node **p;
1092         struct rb_node *parent;
1093         int err = 0;
1094         u64 last_start;
1095         u64 last_end;
1096         bool first_iteration = true;
1097
1098         btrfs_debug_check_extent_io_range(tree, start, end);
1099
1100 again:
1101         if (!prealloc && gfpflags_allow_blocking(mask)) {
1102                 /*
1103                  * Best effort, don't worry if extent state allocation fails
1104                  * here for the first iteration. We might have a cached state
1105                  * that matches exactly the target range, in which case no
1106                  * extent state allocations are needed. We'll only know this
1107                  * after locking the tree.
1108                  */
1109                 prealloc = alloc_extent_state(mask);
1110                 if (!prealloc && !first_iteration)
1111                         return -ENOMEM;
1112         }
1113
1114         spin_lock(&tree->lock);
1115         if (cached_state && *cached_state) {
1116                 state = *cached_state;
1117                 if (state->start <= start && state->end > start &&
1118                     extent_state_in_tree(state)) {
1119                         node = &state->rb_node;
1120                         goto hit_next;
1121                 }
1122         }
1123
1124         /*
1125          * this search will find all the extents that end after
1126          * our range starts.
1127          */
1128         node = tree_search_for_insert(tree, start, &p, &parent);
1129         if (!node) {
1130                 prealloc = alloc_extent_state_atomic(prealloc);
1131                 if (!prealloc) {
1132                         err = -ENOMEM;
1133                         goto out;
1134                 }
1135                 err = insert_state(tree, prealloc, start, end,
1136                                    &p, &parent, &bits, NULL);
1137                 if (err)
1138                         extent_io_tree_panic(tree, err);
1139                 cache_state(prealloc, cached_state);
1140                 prealloc = NULL;
1141                 goto out;
1142         }
1143         state = rb_entry(node, struct extent_state, rb_node);
1144 hit_next:
1145         last_start = state->start;
1146         last_end = state->end;
1147
1148         /*
1149          * | ---- desired range ---- |
1150          * | state |
1151          *
1152          * Just lock what we found and keep going
1153          */
1154         if (state->start == start && state->end <= end) {
1155                 set_state_bits(tree, state, &bits, NULL);
1156                 cache_state(state, cached_state);
1157                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1158                 if (last_end == (u64)-1)
1159                         goto out;
1160                 start = last_end + 1;
1161                 if (start < end && state && state->start == start &&
1162                     !need_resched())
1163                         goto hit_next;
1164                 goto search_again;
1165         }
1166
1167         /*
1168          *     | ---- desired range ---- |
1169          * | state |
1170          *   or
1171          * | ------------- state -------------- |
1172          *
1173          * We need to split the extent we found, and may flip bits on
1174          * second half.
1175          *
1176          * If the extent we found extends past our
1177          * range, we just split and search again.  It'll get split
1178          * again the next time though.
1179          *
1180          * If the extent we found is inside our range, we set the
1181          * desired bit on it.
1182          */
1183         if (state->start < start) {
1184                 prealloc = alloc_extent_state_atomic(prealloc);
1185                 if (!prealloc) {
1186                         err = -ENOMEM;
1187                         goto out;
1188                 }
1189                 err = split_state(tree, state, prealloc, start);
1190                 if (err)
1191                         extent_io_tree_panic(tree, err);
1192                 prealloc = NULL;
1193                 if (err)
1194                         goto out;
1195                 if (state->end <= end) {
1196                         set_state_bits(tree, state, &bits, NULL);
1197                         cache_state(state, cached_state);
1198                         state = clear_state_bit(tree, state, &clear_bits, 0,
1199                                                 NULL);
1200                         if (last_end == (u64)-1)
1201                                 goto out;
1202                         start = last_end + 1;
1203                         if (start < end && state && state->start == start &&
1204                             !need_resched())
1205                                 goto hit_next;
1206                 }
1207                 goto search_again;
1208         }
1209         /*
1210          * | ---- desired range ---- |
1211          *     | state | or               | state |
1212          *
1213          * There's a hole, we need to insert something in it and
1214          * ignore the extent we found.
1215          */
1216         if (state->start > start) {
1217                 u64 this_end;
1218                 if (end < last_start)
1219                         this_end = end;
1220                 else
1221                         this_end = last_start - 1;
1222
1223                 prealloc = alloc_extent_state_atomic(prealloc);
1224                 if (!prealloc) {
1225                         err = -ENOMEM;
1226                         goto out;
1227                 }
1228
1229                 /*
1230                  * Avoid to free 'prealloc' if it can be merged with
1231                  * the later extent.
1232                  */
1233                 err = insert_state(tree, prealloc, start, this_end,
1234                                    NULL, NULL, &bits, NULL);
1235                 if (err)
1236                         extent_io_tree_panic(tree, err);
1237                 cache_state(prealloc, cached_state);
1238                 prealloc = NULL;
1239                 start = this_end + 1;
1240                 goto search_again;
1241         }
1242         /*
1243          * | ---- desired range ---- |
1244          *                        | state |
1245          * We need to split the extent, and set the bit
1246          * on the first half
1247          */
1248         if (state->start <= end && state->end > end) {
1249                 prealloc = alloc_extent_state_atomic(prealloc);
1250                 if (!prealloc) {
1251                         err = -ENOMEM;
1252                         goto out;
1253                 }
1254
1255                 err = split_state(tree, state, prealloc, end + 1);
1256                 if (err)
1257                         extent_io_tree_panic(tree, err);
1258
1259                 set_state_bits(tree, prealloc, &bits, NULL);
1260                 cache_state(prealloc, cached_state);
1261                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1262                 prealloc = NULL;
1263                 goto out;
1264         }
1265
1266         goto search_again;
1267
1268 out:
1269         spin_unlock(&tree->lock);
1270         if (prealloc)
1271                 free_extent_state(prealloc);
1272
1273         return err;
1274
1275 search_again:
1276         if (start > end)
1277                 goto out;
1278         spin_unlock(&tree->lock);
1279         if (gfpflags_allow_blocking(mask))
1280                 cond_resched();
1281         first_iteration = false;
1282         goto again;
1283 }
1284
1285 /* wrappers around set/clear extent bit */
1286 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1287                            unsigned bits, gfp_t mask,
1288                            struct extent_changeset *changeset)
1289 {
1290         /*
1291          * We don't support EXTENT_LOCKED yet, as current changeset will
1292          * record any bits changed, so for EXTENT_LOCKED case, it will
1293          * either fail with -EEXIST or changeset will record the whole
1294          * range.
1295          */
1296         BUG_ON(bits & EXTENT_LOCKED);
1297
1298         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, mask,
1299                                 changeset);
1300 }
1301
1302 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1303                      unsigned bits, int wake, int delete,
1304                      struct extent_state **cached, gfp_t mask)
1305 {
1306         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1307                                   cached, mask, NULL);
1308 }
1309
1310 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1311                              unsigned bits, gfp_t mask,
1312                              struct extent_changeset *changeset)
1313 {
1314         /*
1315          * Don't support EXTENT_LOCKED case, same reason as
1316          * set_record_extent_bits().
1317          */
1318         BUG_ON(bits & EXTENT_LOCKED);
1319
1320         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask,
1321                                   changeset);
1322 }
1323
1324 /*
1325  * either insert or lock state struct between start and end use mask to tell
1326  * us if waiting is desired.
1327  */
1328 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1329                      struct extent_state **cached_state)
1330 {
1331         int err;
1332         u64 failed_start;
1333
1334         while (1) {
1335                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1336                                        EXTENT_LOCKED, &failed_start,
1337                                        cached_state, GFP_NOFS, NULL);
1338                 if (err == -EEXIST) {
1339                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1340                         start = failed_start;
1341                 } else
1342                         break;
1343                 WARN_ON(start > end);
1344         }
1345         return err;
1346 }
1347
1348 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1349 {
1350         int err;
1351         u64 failed_start;
1352
1353         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1354                                &failed_start, NULL, GFP_NOFS, NULL);
1355         if (err == -EEXIST) {
1356                 if (failed_start > start)
1357                         clear_extent_bit(tree, start, failed_start - 1,
1358                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1359                 return 0;
1360         }
1361         return 1;
1362 }
1363
1364 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1365 {
1366         unsigned long index = start >> PAGE_CACHE_SHIFT;
1367         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1368         struct page *page;
1369
1370         while (index <= end_index) {
1371                 page = find_get_page(inode->i_mapping, index);
1372                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1373                 clear_page_dirty_for_io(page);
1374                 page_cache_release(page);
1375                 index++;
1376         }
1377 }
1378
1379 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1380 {
1381         unsigned long index = start >> PAGE_CACHE_SHIFT;
1382         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1383         struct page *page;
1384
1385         while (index <= end_index) {
1386                 page = find_get_page(inode->i_mapping, index);
1387                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1388                 __set_page_dirty_nobuffers(page);
1389                 account_page_redirty(page);
1390                 page_cache_release(page);
1391                 index++;
1392         }
1393 }
1394
1395 /*
1396  * helper function to set both pages and extents in the tree writeback
1397  */
1398 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1399 {
1400         unsigned long index = start >> PAGE_CACHE_SHIFT;
1401         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1402         struct page *page;
1403
1404         while (index <= end_index) {
1405                 page = find_get_page(tree->mapping, index);
1406                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1407                 set_page_writeback(page);
1408                 page_cache_release(page);
1409                 index++;
1410         }
1411 }
1412
1413 /* find the first state struct with 'bits' set after 'start', and
1414  * return it.  tree->lock must be held.  NULL will returned if
1415  * nothing was found after 'start'
1416  */
1417 static struct extent_state *
1418 find_first_extent_bit_state(struct extent_io_tree *tree,
1419                             u64 start, unsigned bits)
1420 {
1421         struct rb_node *node;
1422         struct extent_state *state;
1423
1424         /*
1425          * this search will find all the extents that end after
1426          * our range starts.
1427          */
1428         node = tree_search(tree, start);
1429         if (!node)
1430                 goto out;
1431
1432         while (1) {
1433                 state = rb_entry(node, struct extent_state, rb_node);
1434                 if (state->end >= start && (state->state & bits))
1435                         return state;
1436
1437                 node = rb_next(node);
1438                 if (!node)
1439                         break;
1440         }
1441 out:
1442         return NULL;
1443 }
1444
1445 /*
1446  * find the first offset in the io tree with 'bits' set. zero is
1447  * returned if we find something, and *start_ret and *end_ret are
1448  * set to reflect the state struct that was found.
1449  *
1450  * If nothing was found, 1 is returned. If found something, return 0.
1451  */
1452 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1453                           u64 *start_ret, u64 *end_ret, unsigned bits,
1454                           struct extent_state **cached_state)
1455 {
1456         struct extent_state *state;
1457         struct rb_node *n;
1458         int ret = 1;
1459
1460         spin_lock(&tree->lock);
1461         if (cached_state && *cached_state) {
1462                 state = *cached_state;
1463                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1464                         n = rb_next(&state->rb_node);
1465                         while (n) {
1466                                 state = rb_entry(n, struct extent_state,
1467                                                  rb_node);
1468                                 if (state->state & bits)
1469                                         goto got_it;
1470                                 n = rb_next(n);
1471                         }
1472                         free_extent_state(*cached_state);
1473                         *cached_state = NULL;
1474                         goto out;
1475                 }
1476                 free_extent_state(*cached_state);
1477                 *cached_state = NULL;
1478         }
1479
1480         state = find_first_extent_bit_state(tree, start, bits);
1481 got_it:
1482         if (state) {
1483                 cache_state_if_flags(state, cached_state, 0);
1484                 *start_ret = state->start;
1485                 *end_ret = state->end;
1486                 ret = 0;
1487         }
1488 out:
1489         spin_unlock(&tree->lock);
1490         return ret;
1491 }
1492
1493 /*
1494  * find a contiguous range of bytes in the file marked as delalloc, not
1495  * more than 'max_bytes'.  start and end are used to return the range,
1496  *
1497  * 1 is returned if we find something, 0 if nothing was in the tree
1498  */
1499 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1500                                         u64 *start, u64 *end, u64 max_bytes,
1501                                         struct extent_state **cached_state)
1502 {
1503         struct rb_node *node;
1504         struct extent_state *state;
1505         u64 cur_start = *start;
1506         u64 found = 0;
1507         u64 total_bytes = 0;
1508
1509         spin_lock(&tree->lock);
1510
1511         /*
1512          * this search will find all the extents that end after
1513          * our range starts.
1514          */
1515         node = tree_search(tree, cur_start);
1516         if (!node) {
1517                 if (!found)
1518                         *end = (u64)-1;
1519                 goto out;
1520         }
1521
1522         while (1) {
1523                 state = rb_entry(node, struct extent_state, rb_node);
1524                 if (found && (state->start != cur_start ||
1525                               (state->state & EXTENT_BOUNDARY))) {
1526                         goto out;
1527                 }
1528                 if (!(state->state & EXTENT_DELALLOC)) {
1529                         if (!found)
1530                                 *end = state->end;
1531                         goto out;
1532                 }
1533                 if (!found) {
1534                         *start = state->start;
1535                         *cached_state = state;
1536                         atomic_inc(&state->refs);
1537                 }
1538                 found++;
1539                 *end = state->end;
1540                 cur_start = state->end + 1;
1541                 node = rb_next(node);
1542                 total_bytes += state->end - state->start + 1;
1543                 if (total_bytes >= max_bytes)
1544                         break;
1545                 if (!node)
1546                         break;
1547         }
1548 out:
1549         spin_unlock(&tree->lock);
1550         return found;
1551 }
1552
1553 static noinline void __unlock_for_delalloc(struct inode *inode,
1554                                            struct page *locked_page,
1555                                            u64 start, u64 end)
1556 {
1557         int ret;
1558         struct page *pages[16];
1559         unsigned long index = start >> PAGE_CACHE_SHIFT;
1560         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1561         unsigned long nr_pages = end_index - index + 1;
1562         int i;
1563
1564         if (index == locked_page->index && end_index == index)
1565                 return;
1566
1567         while (nr_pages > 0) {
1568                 ret = find_get_pages_contig(inode->i_mapping, index,
1569                                      min_t(unsigned long, nr_pages,
1570                                      ARRAY_SIZE(pages)), pages);
1571                 for (i = 0; i < ret; i++) {
1572                         if (pages[i] != locked_page)
1573                                 unlock_page(pages[i]);
1574                         page_cache_release(pages[i]);
1575                 }
1576                 nr_pages -= ret;
1577                 index += ret;
1578                 cond_resched();
1579         }
1580 }
1581
1582 static noinline int lock_delalloc_pages(struct inode *inode,
1583                                         struct page *locked_page,
1584                                         u64 delalloc_start,
1585                                         u64 delalloc_end)
1586 {
1587         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1588         unsigned long start_index = index;
1589         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1590         unsigned long pages_locked = 0;
1591         struct page *pages[16];
1592         unsigned long nrpages;
1593         int ret;
1594         int i;
1595
1596         /* the caller is responsible for locking the start index */
1597         if (index == locked_page->index && index == end_index)
1598                 return 0;
1599
1600         /* skip the page at the start index */
1601         nrpages = end_index - index + 1;
1602         while (nrpages > 0) {
1603                 ret = find_get_pages_contig(inode->i_mapping, index,
1604                                      min_t(unsigned long,
1605                                      nrpages, ARRAY_SIZE(pages)), pages);
1606                 if (ret == 0) {
1607                         ret = -EAGAIN;
1608                         goto done;
1609                 }
1610                 /* now we have an array of pages, lock them all */
1611                 for (i = 0; i < ret; i++) {
1612                         /*
1613                          * the caller is taking responsibility for
1614                          * locked_page
1615                          */
1616                         if (pages[i] != locked_page) {
1617                                 lock_page(pages[i]);
1618                                 if (!PageDirty(pages[i]) ||
1619                                     pages[i]->mapping != inode->i_mapping) {
1620                                         ret = -EAGAIN;
1621                                         unlock_page(pages[i]);
1622                                         page_cache_release(pages[i]);
1623                                         goto done;
1624                                 }
1625                         }
1626                         page_cache_release(pages[i]);
1627                         pages_locked++;
1628                 }
1629                 nrpages -= ret;
1630                 index += ret;
1631                 cond_resched();
1632         }
1633         ret = 0;
1634 done:
1635         if (ret && pages_locked) {
1636                 __unlock_for_delalloc(inode, locked_page,
1637                               delalloc_start,
1638                               ((u64)(start_index + pages_locked - 1)) <<
1639                               PAGE_CACHE_SHIFT);
1640         }
1641         return ret;
1642 }
1643
1644 /*
1645  * find a contiguous range of bytes in the file marked as delalloc, not
1646  * more than 'max_bytes'.  start and end are used to return the range,
1647  *
1648  * 1 is returned if we find something, 0 if nothing was in the tree
1649  */
1650 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1651                                     struct extent_io_tree *tree,
1652                                     struct page *locked_page, u64 *start,
1653                                     u64 *end, u64 max_bytes)
1654 {
1655         u64 delalloc_start;
1656         u64 delalloc_end;
1657         u64 found;
1658         struct extent_state *cached_state = NULL;
1659         int ret;
1660         int loops = 0;
1661
1662 again:
1663         /* step one, find a bunch of delalloc bytes starting at start */
1664         delalloc_start = *start;
1665         delalloc_end = 0;
1666         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1667                                     max_bytes, &cached_state);
1668         if (!found || delalloc_end <= *start) {
1669                 *start = delalloc_start;
1670                 *end = delalloc_end;
1671                 free_extent_state(cached_state);
1672                 return 0;
1673         }
1674
1675         /*
1676          * start comes from the offset of locked_page.  We have to lock
1677          * pages in order, so we can't process delalloc bytes before
1678          * locked_page
1679          */
1680         if (delalloc_start < *start)
1681                 delalloc_start = *start;
1682
1683         /*
1684          * make sure to limit the number of pages we try to lock down
1685          */
1686         if (delalloc_end + 1 - delalloc_start > max_bytes)
1687                 delalloc_end = delalloc_start + max_bytes - 1;
1688
1689         /* step two, lock all the pages after the page that has start */
1690         ret = lock_delalloc_pages(inode, locked_page,
1691                                   delalloc_start, delalloc_end);
1692         if (ret == -EAGAIN) {
1693                 /* some of the pages are gone, lets avoid looping by
1694                  * shortening the size of the delalloc range we're searching
1695                  */
1696                 free_extent_state(cached_state);
1697                 cached_state = NULL;
1698                 if (!loops) {
1699                         max_bytes = PAGE_CACHE_SIZE;
1700                         loops = 1;
1701                         goto again;
1702                 } else {
1703                         found = 0;
1704                         goto out_failed;
1705                 }
1706         }
1707         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1708
1709         /* step three, lock the state bits for the whole range */
1710         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1711
1712         /* then test to make sure it is all still delalloc */
1713         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1714                              EXTENT_DELALLOC, 1, cached_state);
1715         if (!ret) {
1716                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1717                                      &cached_state, GFP_NOFS);
1718                 __unlock_for_delalloc(inode, locked_page,
1719                               delalloc_start, delalloc_end);
1720                 cond_resched();
1721                 goto again;
1722         }
1723         free_extent_state(cached_state);
1724         *start = delalloc_start;
1725         *end = delalloc_end;
1726 out_failed:
1727         return found;
1728 }
1729
1730 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1731                                  struct page *locked_page,
1732                                  unsigned clear_bits,
1733                                  unsigned long page_ops)
1734 {
1735         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1736         int ret;
1737         struct page *pages[16];
1738         unsigned long index = start >> PAGE_CACHE_SHIFT;
1739         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1740         unsigned long nr_pages = end_index - index + 1;
1741         int i;
1742
1743         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1744         if (page_ops == 0)
1745                 return;
1746
1747         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1748                 mapping_set_error(inode->i_mapping, -EIO);
1749
1750         while (nr_pages > 0) {
1751                 ret = find_get_pages_contig(inode->i_mapping, index,
1752                                      min_t(unsigned long,
1753                                      nr_pages, ARRAY_SIZE(pages)), pages);
1754                 for (i = 0; i < ret; i++) {
1755
1756                         if (page_ops & PAGE_SET_PRIVATE2)
1757                                 SetPagePrivate2(pages[i]);
1758
1759                         if (pages[i] == locked_page) {
1760                                 page_cache_release(pages[i]);
1761                                 continue;
1762                         }
1763                         if (page_ops & PAGE_CLEAR_DIRTY)
1764                                 clear_page_dirty_for_io(pages[i]);
1765                         if (page_ops & PAGE_SET_WRITEBACK)
1766                                 set_page_writeback(pages[i]);
1767                         if (page_ops & PAGE_SET_ERROR)
1768                                 SetPageError(pages[i]);
1769                         if (page_ops & PAGE_END_WRITEBACK)
1770                                 end_page_writeback(pages[i]);
1771                         if (page_ops & PAGE_UNLOCK)
1772                                 unlock_page(pages[i]);
1773                         page_cache_release(pages[i]);
1774                 }
1775                 nr_pages -= ret;
1776                 index += ret;
1777                 cond_resched();
1778         }
1779 }
1780
1781 /*
1782  * count the number of bytes in the tree that have a given bit(s)
1783  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1784  * cached.  The total number found is returned.
1785  */
1786 u64 count_range_bits(struct extent_io_tree *tree,
1787                      u64 *start, u64 search_end, u64 max_bytes,
1788                      unsigned bits, int contig)
1789 {
1790         struct rb_node *node;
1791         struct extent_state *state;
1792         u64 cur_start = *start;
1793         u64 total_bytes = 0;
1794         u64 last = 0;
1795         int found = 0;
1796
1797         if (WARN_ON(search_end <= cur_start))
1798                 return 0;
1799
1800         spin_lock(&tree->lock);
1801         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1802                 total_bytes = tree->dirty_bytes;
1803                 goto out;
1804         }
1805         /*
1806          * this search will find all the extents that end after
1807          * our range starts.
1808          */
1809         node = tree_search(tree, cur_start);
1810         if (!node)
1811                 goto out;
1812
1813         while (1) {
1814                 state = rb_entry(node, struct extent_state, rb_node);
1815                 if (state->start > search_end)
1816                         break;
1817                 if (contig && found && state->start > last + 1)
1818                         break;
1819                 if (state->end >= cur_start && (state->state & bits) == bits) {
1820                         total_bytes += min(search_end, state->end) + 1 -
1821                                        max(cur_start, state->start);
1822                         if (total_bytes >= max_bytes)
1823                                 break;
1824                         if (!found) {
1825                                 *start = max(cur_start, state->start);
1826                                 found = 1;
1827                         }
1828                         last = state->end;
1829                 } else if (contig && found) {
1830                         break;
1831                 }
1832                 node = rb_next(node);
1833                 if (!node)
1834                         break;
1835         }
1836 out:
1837         spin_unlock(&tree->lock);
1838         return total_bytes;
1839 }
1840
1841 /*
1842  * set the private field for a given byte offset in the tree.  If there isn't
1843  * an extent_state there already, this does nothing.
1844  */
1845 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1846                 struct io_failure_record *failrec)
1847 {
1848         struct rb_node *node;
1849         struct extent_state *state;
1850         int ret = 0;
1851
1852         spin_lock(&tree->lock);
1853         /*
1854          * this search will find all the extents that end after
1855          * our range starts.
1856          */
1857         node = tree_search(tree, start);
1858         if (!node) {
1859                 ret = -ENOENT;
1860                 goto out;
1861         }
1862         state = rb_entry(node, struct extent_state, rb_node);
1863         if (state->start != start) {
1864                 ret = -ENOENT;
1865                 goto out;
1866         }
1867         state->failrec = failrec;
1868 out:
1869         spin_unlock(&tree->lock);
1870         return ret;
1871 }
1872
1873 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1874                 struct io_failure_record **failrec)
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         *failrec = state->failrec;
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_failrec(failure_tree, rec->start, NULL);
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         struct io_failure_record *failrec;
2093         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2094         struct extent_state *state;
2095         int num_copies;
2096         int ret;
2097
2098         private = 0;
2099         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2100                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2101         if (!ret)
2102                 return 0;
2103
2104         ret = get_state_failrec(&BTRFS_I(inode)->io_failure_tree, start,
2105                         &failrec);
2106         if (ret)
2107                 return 0;
2108
2109         BUG_ON(!failrec->this_mirror);
2110
2111         if (failrec->in_validation) {
2112                 /* there was no real error, just free the record */
2113                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2114                          failrec->start);
2115                 goto out;
2116         }
2117         if (fs_info->sb->s_flags & MS_RDONLY)
2118                 goto out;
2119
2120         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2121         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2122                                             failrec->start,
2123                                             EXTENT_LOCKED);
2124         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2125
2126         if (state && state->start <= failrec->start &&
2127             state->end >= failrec->start + failrec->len - 1) {
2128                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2129                                               failrec->len);
2130                 if (num_copies > 1)  {
2131                         repair_io_failure(inode, start, failrec->len,
2132                                           failrec->logical, page,
2133                                           pg_offset, failrec->failed_mirror);
2134                 }
2135         }
2136
2137 out:
2138         free_io_failure(inode, failrec);
2139
2140         return 0;
2141 }
2142
2143 /*
2144  * Can be called when
2145  * - hold extent lock
2146  * - under ordered extent
2147  * - the inode is freeing
2148  */
2149 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2150 {
2151         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2152         struct io_failure_record *failrec;
2153         struct extent_state *state, *next;
2154
2155         if (RB_EMPTY_ROOT(&failure_tree->state))
2156                 return;
2157
2158         spin_lock(&failure_tree->lock);
2159         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2160         while (state) {
2161                 if (state->start > end)
2162                         break;
2163
2164                 ASSERT(state->end <= end);
2165
2166                 next = next_state(state);
2167
2168                 failrec = state->failrec;
2169                 free_extent_state(state);
2170                 kfree(failrec);
2171
2172                 state = next;
2173         }
2174         spin_unlock(&failure_tree->lock);
2175 }
2176
2177 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2178                 struct io_failure_record **failrec_ret)
2179 {
2180         struct io_failure_record *failrec;
2181         struct extent_map *em;
2182         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2183         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2184         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2185         int ret;
2186         u64 logical;
2187
2188         ret = get_state_failrec(failure_tree, start, &failrec);
2189         if (ret) {
2190                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2191                 if (!failrec)
2192                         return -ENOMEM;
2193
2194                 failrec->start = start;
2195                 failrec->len = end - start + 1;
2196                 failrec->this_mirror = 0;
2197                 failrec->bio_flags = 0;
2198                 failrec->in_validation = 0;
2199
2200                 read_lock(&em_tree->lock);
2201                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2202                 if (!em) {
2203                         read_unlock(&em_tree->lock);
2204                         kfree(failrec);
2205                         return -EIO;
2206                 }
2207
2208                 if (em->start > start || em->start + em->len <= start) {
2209                         free_extent_map(em);
2210                         em = NULL;
2211                 }
2212                 read_unlock(&em_tree->lock);
2213                 if (!em) {
2214                         kfree(failrec);
2215                         return -EIO;
2216                 }
2217
2218                 logical = start - em->start;
2219                 logical = em->block_start + logical;
2220                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2221                         logical = em->block_start;
2222                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2223                         extent_set_compress_type(&failrec->bio_flags,
2224                                                  em->compress_type);
2225                 }
2226
2227                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2228                          logical, start, failrec->len);
2229
2230                 failrec->logical = logical;
2231                 free_extent_map(em);
2232
2233                 /* set the bits in the private failure tree */
2234                 ret = set_extent_bits(failure_tree, start, end,
2235                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2236                 if (ret >= 0)
2237                         ret = set_state_failrec(failure_tree, start, failrec);
2238                 /* set the bits in the inode's tree */
2239                 if (ret >= 0)
2240                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2241                                                 GFP_NOFS);
2242                 if (ret < 0) {
2243                         kfree(failrec);
2244                         return ret;
2245                 }
2246         } else {
2247                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2248                          failrec->logical, failrec->start, failrec->len,
2249                          failrec->in_validation);
2250                 /*
2251                  * when data can be on disk more than twice, add to failrec here
2252                  * (e.g. with a list for failed_mirror) to make
2253                  * clean_io_failure() clean all those errors at once.
2254                  */
2255         }
2256
2257         *failrec_ret = failrec;
2258
2259         return 0;
2260 }
2261
2262 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2263                            struct io_failure_record *failrec, int failed_mirror)
2264 {
2265         int num_copies;
2266
2267         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2268                                       failrec->logical, failrec->len);
2269         if (num_copies == 1) {
2270                 /*
2271                  * we only have a single copy of the data, so don't bother with
2272                  * all the retry and error correction code that follows. no
2273                  * matter what the error is, it is very likely to persist.
2274                  */
2275                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2276                          num_copies, failrec->this_mirror, failed_mirror);
2277                 return 0;
2278         }
2279
2280         /*
2281          * there are two premises:
2282          *      a) deliver good data to the caller
2283          *      b) correct the bad sectors on disk
2284          */
2285         if (failed_bio->bi_vcnt > 1) {
2286                 /*
2287                  * to fulfill b), we need to know the exact failing sectors, as
2288                  * we don't want to rewrite any more than the failed ones. thus,
2289                  * we need separate read requests for the failed bio
2290                  *
2291                  * if the following BUG_ON triggers, our validation request got
2292                  * merged. we need separate requests for our algorithm to work.
2293                  */
2294                 BUG_ON(failrec->in_validation);
2295                 failrec->in_validation = 1;
2296                 failrec->this_mirror = failed_mirror;
2297         } else {
2298                 /*
2299                  * we're ready to fulfill a) and b) alongside. get a good copy
2300                  * of the failed sector and if we succeed, we have setup
2301                  * everything for repair_io_failure to do the rest for us.
2302                  */
2303                 if (failrec->in_validation) {
2304                         BUG_ON(failrec->this_mirror != failed_mirror);
2305                         failrec->in_validation = 0;
2306                         failrec->this_mirror = 0;
2307                 }
2308                 failrec->failed_mirror = failed_mirror;
2309                 failrec->this_mirror++;
2310                 if (failrec->this_mirror == failed_mirror)
2311                         failrec->this_mirror++;
2312         }
2313
2314         if (failrec->this_mirror > num_copies) {
2315                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2316                          num_copies, failrec->this_mirror, failed_mirror);
2317                 return 0;
2318         }
2319
2320         return 1;
2321 }
2322
2323
2324 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2325                                     struct io_failure_record *failrec,
2326                                     struct page *page, int pg_offset, int icsum,
2327                                     bio_end_io_t *endio_func, void *data)
2328 {
2329         struct bio *bio;
2330         struct btrfs_io_bio *btrfs_failed_bio;
2331         struct btrfs_io_bio *btrfs_bio;
2332
2333         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2334         if (!bio)
2335                 return NULL;
2336
2337         bio->bi_end_io = endio_func;
2338         bio->bi_iter.bi_sector = failrec->logical >> 9;
2339         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2340         bio->bi_iter.bi_size = 0;
2341         bio->bi_private = data;
2342
2343         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2344         if (btrfs_failed_bio->csum) {
2345                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2346                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2347
2348                 btrfs_bio = btrfs_io_bio(bio);
2349                 btrfs_bio->csum = btrfs_bio->csum_inline;
2350                 icsum *= csum_size;
2351                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2352                        csum_size);
2353         }
2354
2355         bio_add_page(bio, page, failrec->len, pg_offset);
2356
2357         return bio;
2358 }
2359
2360 /*
2361  * this is a generic handler for readpage errors (default
2362  * readpage_io_failed_hook). if other copies exist, read those and write back
2363  * good data to the failed position. does not investigate in remapping the
2364  * failed extent elsewhere, hoping the device will be smart enough to do this as
2365  * needed
2366  */
2367
2368 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2369                               struct page *page, u64 start, u64 end,
2370                               int failed_mirror)
2371 {
2372         struct io_failure_record *failrec;
2373         struct inode *inode = page->mapping->host;
2374         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2375         struct bio *bio;
2376         int read_mode;
2377         int ret;
2378
2379         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2380
2381         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2382         if (ret)
2383                 return ret;
2384
2385         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2386         if (!ret) {
2387                 free_io_failure(inode, failrec);
2388                 return -EIO;
2389         }
2390
2391         if (failed_bio->bi_vcnt > 1)
2392                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2393         else
2394                 read_mode = READ_SYNC;
2395
2396         phy_offset >>= inode->i_sb->s_blocksize_bits;
2397         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2398                                       start - page_offset(page),
2399                                       (int)phy_offset, failed_bio->bi_end_io,
2400                                       NULL);
2401         if (!bio) {
2402                 free_io_failure(inode, failrec);
2403                 return -EIO;
2404         }
2405
2406         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2407                  read_mode, failrec->this_mirror, failrec->in_validation);
2408
2409         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2410                                          failrec->this_mirror,
2411                                          failrec->bio_flags, 0);
2412         if (ret) {
2413                 free_io_failure(inode, failrec);
2414                 bio_put(bio);
2415         }
2416
2417         return ret;
2418 }
2419
2420 /* lots and lots of room for performance fixes in the end_bio funcs */
2421
2422 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2423 {
2424         int uptodate = (err == 0);
2425         struct extent_io_tree *tree;
2426         int ret = 0;
2427
2428         tree = &BTRFS_I(page->mapping->host)->io_tree;
2429
2430         if (tree->ops && tree->ops->writepage_end_io_hook) {
2431                 ret = tree->ops->writepage_end_io_hook(page, start,
2432                                                end, NULL, uptodate);
2433                 if (ret)
2434                         uptodate = 0;
2435         }
2436
2437         if (!uptodate) {
2438                 ClearPageUptodate(page);
2439                 SetPageError(page);
2440                 ret = ret < 0 ? ret : -EIO;
2441                 mapping_set_error(page->mapping, ret);
2442         }
2443 }
2444
2445 /*
2446  * after a writepage IO is done, we need to:
2447  * clear the uptodate bits on error
2448  * clear the writeback bits in the extent tree for this IO
2449  * end_page_writeback if the page has no more pending IO
2450  *
2451  * Scheduling is not allowed, so the extent state tree is expected
2452  * to have one and only one object corresponding to this IO.
2453  */
2454 static void end_bio_extent_writepage(struct bio *bio)
2455 {
2456         struct bio_vec *bvec;
2457         u64 start;
2458         u64 end;
2459         int i;
2460
2461         bio_for_each_segment_all(bvec, bio, i) {
2462                 struct page *page = bvec->bv_page;
2463
2464                 /* We always issue full-page reads, but if some block
2465                  * in a page fails to read, blk_update_request() will
2466                  * advance bv_offset and adjust bv_len to compensate.
2467                  * Print a warning for nonzero offsets, and an error
2468                  * if they don't add up to a full page.  */
2469                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2470                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2471                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2472                                    "partial page write in btrfs with offset %u and length %u",
2473                                         bvec->bv_offset, bvec->bv_len);
2474                         else
2475                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2476                                    "incomplete page write in btrfs with offset %u and "
2477                                    "length %u",
2478                                         bvec->bv_offset, bvec->bv_len);
2479                 }
2480
2481                 start = page_offset(page);
2482                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2483
2484                 end_extent_writepage(page, bio->bi_error, start, end);
2485                 end_page_writeback(page);
2486         }
2487
2488         bio_put(bio);
2489 }
2490
2491 static void
2492 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2493                               int uptodate)
2494 {
2495         struct extent_state *cached = NULL;
2496         u64 end = start + len - 1;
2497
2498         if (uptodate && tree->track_uptodate)
2499                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2500         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2501 }
2502
2503 /*
2504  * after a readpage IO is done, we need to:
2505  * clear the uptodate bits on error
2506  * set the uptodate bits if things worked
2507  * set the page up to date if all extents in the tree are uptodate
2508  * clear the lock bit in the extent tree
2509  * unlock the page if there are no other extents locked for it
2510  *
2511  * Scheduling is not allowed, so the extent state tree is expected
2512  * to have one and only one object corresponding to this IO.
2513  */
2514 static void end_bio_extent_readpage(struct bio *bio)
2515 {
2516         struct bio_vec *bvec;
2517         int uptodate = !bio->bi_error;
2518         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2519         struct extent_io_tree *tree;
2520         u64 offset = 0;
2521         u64 start;
2522         u64 end;
2523         u64 len;
2524         u64 extent_start = 0;
2525         u64 extent_len = 0;
2526         int mirror;
2527         int ret;
2528         int i;
2529
2530         bio_for_each_segment_all(bvec, bio, i) {
2531                 struct page *page = bvec->bv_page;
2532                 struct inode *inode = page->mapping->host;
2533
2534                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2535                          "mirror=%u\n", (u64)bio->bi_iter.bi_sector,
2536                          bio->bi_error, io_bio->mirror_num);
2537                 tree = &BTRFS_I(inode)->io_tree;
2538
2539                 /* We always issue full-page reads, but if some block
2540                  * in a page fails to read, blk_update_request() will
2541                  * advance bv_offset and adjust bv_len to compensate.
2542                  * Print a warning for nonzero offsets, and an error
2543                  * if they don't add up to a full page.  */
2544                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2545                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2546                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2547                                    "partial page read in btrfs with offset %u and length %u",
2548                                         bvec->bv_offset, bvec->bv_len);
2549                         else
2550                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2551                                    "incomplete page read in btrfs with offset %u and "
2552                                    "length %u",
2553                                         bvec->bv_offset, bvec->bv_len);
2554                 }
2555
2556                 start = page_offset(page);
2557                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2558                 len = bvec->bv_len;
2559
2560                 mirror = io_bio->mirror_num;
2561                 if (likely(uptodate && tree->ops &&
2562                            tree->ops->readpage_end_io_hook)) {
2563                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2564                                                               page, start, end,
2565                                                               mirror);
2566                         if (ret)
2567                                 uptodate = 0;
2568                         else
2569                                 clean_io_failure(inode, start, page, 0);
2570                 }
2571
2572                 if (likely(uptodate))
2573                         goto readpage_ok;
2574
2575                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2576                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2577                         if (!ret && !bio->bi_error)
2578                                 uptodate = 1;
2579                 } else {
2580                         /*
2581                          * The generic bio_readpage_error handles errors the
2582                          * following way: If possible, new read requests are
2583                          * created and submitted and will end up in
2584                          * end_bio_extent_readpage as well (if we're lucky, not
2585                          * in the !uptodate case). In that case it returns 0 and
2586                          * we just go on with the next page in our bio. If it
2587                          * can't handle the error it will return -EIO and we
2588                          * remain responsible for that page.
2589                          */
2590                         ret = bio_readpage_error(bio, offset, page, start, end,
2591                                                  mirror);
2592                         if (ret == 0) {
2593                                 uptodate = !bio->bi_error;
2594                                 offset += len;
2595                                 continue;
2596                         }
2597                 }
2598 readpage_ok:
2599                 if (likely(uptodate)) {
2600                         loff_t i_size = i_size_read(inode);
2601                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2602                         unsigned off;
2603
2604                         /* Zero out the end if this page straddles i_size */
2605                         off = i_size & (PAGE_CACHE_SIZE-1);
2606                         if (page->index == end_index && off)
2607                                 zero_user_segment(page, off, PAGE_CACHE_SIZE);
2608                         SetPageUptodate(page);
2609                 } else {
2610                         ClearPageUptodate(page);
2611                         SetPageError(page);
2612                 }
2613                 unlock_page(page);
2614                 offset += len;
2615
2616                 if (unlikely(!uptodate)) {
2617                         if (extent_len) {
2618                                 endio_readpage_release_extent(tree,
2619                                                               extent_start,
2620                                                               extent_len, 1);
2621                                 extent_start = 0;
2622                                 extent_len = 0;
2623                         }
2624                         endio_readpage_release_extent(tree, start,
2625                                                       end - start + 1, 0);
2626                 } else if (!extent_len) {
2627                         extent_start = start;
2628                         extent_len = end + 1 - start;
2629                 } else if (extent_start + extent_len == start) {
2630                         extent_len += end + 1 - start;
2631                 } else {
2632                         endio_readpage_release_extent(tree, extent_start,
2633                                                       extent_len, uptodate);
2634                         extent_start = start;
2635                         extent_len = end + 1 - start;
2636                 }
2637         }
2638
2639         if (extent_len)
2640                 endio_readpage_release_extent(tree, extent_start, extent_len,
2641                                               uptodate);
2642         if (io_bio->end_io)
2643                 io_bio->end_io(io_bio, bio->bi_error);
2644         bio_put(bio);
2645 }
2646
2647 /*
2648  * this allocates from the btrfs_bioset.  We're returning a bio right now
2649  * but you can call btrfs_io_bio for the appropriate container_of magic
2650  */
2651 struct bio *
2652 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2653                 gfp_t gfp_flags)
2654 {
2655         struct btrfs_io_bio *btrfs_bio;
2656         struct bio *bio;
2657
2658         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2659
2660         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2661                 while (!bio && (nr_vecs /= 2)) {
2662                         bio = bio_alloc_bioset(gfp_flags,
2663                                                nr_vecs, btrfs_bioset);
2664                 }
2665         }
2666
2667         if (bio) {
2668                 bio->bi_bdev = bdev;
2669                 bio->bi_iter.bi_sector = first_sector;
2670                 btrfs_bio = btrfs_io_bio(bio);
2671                 btrfs_bio->csum = NULL;
2672                 btrfs_bio->csum_allocated = NULL;
2673                 btrfs_bio->end_io = NULL;
2674         }
2675         return bio;
2676 }
2677
2678 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2679 {
2680         struct btrfs_io_bio *btrfs_bio;
2681         struct bio *new;
2682
2683         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2684         if (new) {
2685                 btrfs_bio = btrfs_io_bio(new);
2686                 btrfs_bio->csum = NULL;
2687                 btrfs_bio->csum_allocated = NULL;
2688                 btrfs_bio->end_io = NULL;
2689
2690 #ifdef CONFIG_BLK_CGROUP
2691                 /* FIXME, put this into bio_clone_bioset */
2692                 if (bio->bi_css)
2693                         bio_associate_blkcg(new, bio->bi_css);
2694 #endif
2695         }
2696         return new;
2697 }
2698
2699 /* this also allocates from the btrfs_bioset */
2700 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2701 {
2702         struct btrfs_io_bio *btrfs_bio;
2703         struct bio *bio;
2704
2705         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2706         if (bio) {
2707                 btrfs_bio = btrfs_io_bio(bio);
2708                 btrfs_bio->csum = NULL;
2709                 btrfs_bio->csum_allocated = NULL;
2710                 btrfs_bio->end_io = NULL;
2711         }
2712         return bio;
2713 }
2714
2715
2716 static int __must_check submit_one_bio(int rw, struct bio *bio,
2717                                        int mirror_num, unsigned long bio_flags)
2718 {
2719         int ret = 0;
2720         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2721         struct page *page = bvec->bv_page;
2722         struct extent_io_tree *tree = bio->bi_private;
2723         u64 start;
2724
2725         start = page_offset(page) + bvec->bv_offset;
2726
2727         bio->bi_private = NULL;
2728
2729         bio_get(bio);
2730
2731         if (tree->ops && tree->ops->submit_bio_hook)
2732                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2733                                            mirror_num, bio_flags, start);
2734         else
2735                 btrfsic_submit_bio(rw, bio);
2736
2737         bio_put(bio);
2738         return ret;
2739 }
2740
2741 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2742                      unsigned long offset, size_t size, struct bio *bio,
2743                      unsigned long bio_flags)
2744 {
2745         int ret = 0;
2746         if (tree->ops && tree->ops->merge_bio_hook)
2747                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2748                                                 bio_flags);
2749         BUG_ON(ret < 0);
2750         return ret;
2751
2752 }
2753
2754 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2755                               struct writeback_control *wbc,
2756                               struct page *page, sector_t sector,
2757                               size_t size, unsigned long offset,
2758                               struct block_device *bdev,
2759                               struct bio **bio_ret,
2760                               unsigned long max_pages,
2761                               bio_end_io_t end_io_func,
2762                               int mirror_num,
2763                               unsigned long prev_bio_flags,
2764                               unsigned long bio_flags,
2765                               bool force_bio_submit)
2766 {
2767         int ret = 0;
2768         struct bio *bio;
2769         int contig = 0;
2770         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2771         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2772
2773         if (bio_ret && *bio_ret) {
2774                 bio = *bio_ret;
2775                 if (old_compressed)
2776                         contig = bio->bi_iter.bi_sector == sector;
2777                 else
2778                         contig = bio_end_sector(bio) == sector;
2779
2780                 if (prev_bio_flags != bio_flags || !contig ||
2781                     force_bio_submit ||
2782                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2783                     bio_add_page(bio, page, page_size, offset) < page_size) {
2784                         ret = submit_one_bio(rw, bio, mirror_num,
2785                                              prev_bio_flags);
2786                         if (ret < 0) {
2787                                 *bio_ret = NULL;
2788                                 return ret;
2789                         }
2790                         bio = NULL;
2791                 } else {
2792                         if (wbc)
2793                                 wbc_account_io(wbc, page, page_size);
2794                         return 0;
2795                 }
2796         }
2797
2798         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2799                         GFP_NOFS | __GFP_HIGH);
2800         if (!bio)
2801                 return -ENOMEM;
2802
2803         bio_add_page(bio, page, page_size, offset);
2804         bio->bi_end_io = end_io_func;
2805         bio->bi_private = tree;
2806         if (wbc) {
2807                 wbc_init_bio(wbc, bio);
2808                 wbc_account_io(wbc, page, page_size);
2809         }
2810
2811         if (bio_ret)
2812                 *bio_ret = bio;
2813         else
2814                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2815
2816         return ret;
2817 }
2818
2819 static void attach_extent_buffer_page(struct extent_buffer *eb,
2820                                       struct page *page)
2821 {
2822         if (!PagePrivate(page)) {
2823                 SetPagePrivate(page);
2824                 page_cache_get(page);
2825                 set_page_private(page, (unsigned long)eb);
2826         } else {
2827                 WARN_ON(page->private != (unsigned long)eb);
2828         }
2829 }
2830
2831 void set_page_extent_mapped(struct page *page)
2832 {
2833         if (!PagePrivate(page)) {
2834                 SetPagePrivate(page);
2835                 page_cache_get(page);
2836                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2837         }
2838 }
2839
2840 static struct extent_map *
2841 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2842                  u64 start, u64 len, get_extent_t *get_extent,
2843                  struct extent_map **em_cached)
2844 {
2845         struct extent_map *em;
2846
2847         if (em_cached && *em_cached) {
2848                 em = *em_cached;
2849                 if (extent_map_in_tree(em) && start >= em->start &&
2850                     start < extent_map_end(em)) {
2851                         atomic_inc(&em->refs);
2852                         return em;
2853                 }
2854
2855                 free_extent_map(em);
2856                 *em_cached = NULL;
2857         }
2858
2859         em = get_extent(inode, page, pg_offset, start, len, 0);
2860         if (em_cached && !IS_ERR_OR_NULL(em)) {
2861                 BUG_ON(*em_cached);
2862                 atomic_inc(&em->refs);
2863                 *em_cached = em;
2864         }
2865         return em;
2866 }
2867 /*
2868  * basic readpage implementation.  Locked extent state structs are inserted
2869  * into the tree that are removed when the IO is done (by the end_io
2870  * handlers)
2871  * XXX JDM: This needs looking at to ensure proper page locking
2872  */
2873 static int __do_readpage(struct extent_io_tree *tree,
2874                          struct page *page,
2875                          get_extent_t *get_extent,
2876                          struct extent_map **em_cached,
2877                          struct bio **bio, int mirror_num,
2878                          unsigned long *bio_flags, int rw,
2879                          u64 *prev_em_start)
2880 {
2881         struct inode *inode = page->mapping->host;
2882         u64 start = page_offset(page);
2883         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2884         u64 end;
2885         u64 cur = start;
2886         u64 extent_offset;
2887         u64 last_byte = i_size_read(inode);
2888         u64 block_start;
2889         u64 cur_end;
2890         sector_t sector;
2891         struct extent_map *em;
2892         struct block_device *bdev;
2893         int ret;
2894         int nr = 0;
2895         size_t pg_offset = 0;
2896         size_t iosize;
2897         size_t disk_io_size;
2898         size_t blocksize = inode->i_sb->s_blocksize;
2899         unsigned long this_bio_flag = 0;
2900
2901         set_page_extent_mapped(page);
2902
2903         end = page_end;
2904         if (!PageUptodate(page)) {
2905                 if (cleancache_get_page(page) == 0) {
2906                         BUG_ON(blocksize != PAGE_SIZE);
2907                         unlock_extent(tree, start, end);
2908                         goto out;
2909                 }
2910         }
2911
2912         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2913                 char *userpage;
2914                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2915
2916                 if (zero_offset) {
2917                         iosize = PAGE_CACHE_SIZE - zero_offset;
2918                         userpage = kmap_atomic(page);
2919                         memset(userpage + zero_offset, 0, iosize);
2920                         flush_dcache_page(page);
2921                         kunmap_atomic(userpage);
2922                 }
2923         }
2924         while (cur <= end) {
2925                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2926                 bool force_bio_submit = false;
2927
2928                 if (cur >= last_byte) {
2929                         char *userpage;
2930                         struct extent_state *cached = NULL;
2931
2932                         iosize = PAGE_CACHE_SIZE - pg_offset;
2933                         userpage = kmap_atomic(page);
2934                         memset(userpage + pg_offset, 0, iosize);
2935                         flush_dcache_page(page);
2936                         kunmap_atomic(userpage);
2937                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2938                                             &cached, GFP_NOFS);
2939                         unlock_extent_cached(tree, cur,
2940                                              cur + iosize - 1,
2941                                              &cached, GFP_NOFS);
2942                         break;
2943                 }
2944                 em = __get_extent_map(inode, page, pg_offset, cur,
2945                                       end - cur + 1, get_extent, em_cached);
2946                 if (IS_ERR_OR_NULL(em)) {
2947                         SetPageError(page);
2948                         unlock_extent(tree, cur, end);
2949                         break;
2950                 }
2951                 extent_offset = cur - em->start;
2952                 BUG_ON(extent_map_end(em) <= cur);
2953                 BUG_ON(end < cur);
2954
2955                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2956                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2957                         extent_set_compress_type(&this_bio_flag,
2958                                                  em->compress_type);
2959                 }
2960
2961                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2962                 cur_end = min(extent_map_end(em) - 1, end);
2963                 iosize = ALIGN(iosize, blocksize);
2964                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2965                         disk_io_size = em->block_len;
2966                         sector = em->block_start >> 9;
2967                 } else {
2968                         sector = (em->block_start + extent_offset) >> 9;
2969                         disk_io_size = iosize;
2970                 }
2971                 bdev = em->bdev;
2972                 block_start = em->block_start;
2973                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2974                         block_start = EXTENT_MAP_HOLE;
2975
2976                 /*
2977                  * If we have a file range that points to a compressed extent
2978                  * and it's followed by a consecutive file range that points to
2979                  * to the same compressed extent (possibly with a different
2980                  * offset and/or length, so it either points to the whole extent
2981                  * or only part of it), we must make sure we do not submit a
2982                  * single bio to populate the pages for the 2 ranges because
2983                  * this makes the compressed extent read zero out the pages
2984                  * belonging to the 2nd range. Imagine the following scenario:
2985                  *
2986                  *  File layout
2987                  *  [0 - 8K]                     [8K - 24K]
2988                  *    |                               |
2989                  *    |                               |
2990                  * points to extent X,         points to extent X,
2991                  * offset 4K, length of 8K     offset 0, length 16K
2992                  *
2993                  * [extent X, compressed length = 4K uncompressed length = 16K]
2994                  *
2995                  * If the bio to read the compressed extent covers both ranges,
2996                  * it will decompress extent X into the pages belonging to the
2997                  * first range and then it will stop, zeroing out the remaining
2998                  * pages that belong to the other range that points to extent X.
2999                  * So here we make sure we submit 2 bios, one for the first
3000                  * range and another one for the third range. Both will target
3001                  * the same physical extent from disk, but we can't currently
3002                  * make the compressed bio endio callback populate the pages
3003                  * for both ranges because each compressed bio is tightly
3004                  * coupled with a single extent map, and each range can have
3005                  * an extent map with a different offset value relative to the
3006                  * uncompressed data of our extent and different lengths. This
3007                  * is a corner case so we prioritize correctness over
3008                  * non-optimal behavior (submitting 2 bios for the same extent).
3009                  */
3010                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3011                     prev_em_start && *prev_em_start != (u64)-1 &&
3012                     *prev_em_start != em->orig_start)
3013                         force_bio_submit = true;
3014
3015                 if (prev_em_start)
3016                         *prev_em_start = em->orig_start;
3017
3018                 free_extent_map(em);
3019                 em = NULL;
3020
3021                 /* we've found a hole, just zero and go on */
3022                 if (block_start == EXTENT_MAP_HOLE) {
3023                         char *userpage;
3024                         struct extent_state *cached = NULL;
3025
3026                         userpage = kmap_atomic(page);
3027                         memset(userpage + pg_offset, 0, iosize);
3028                         flush_dcache_page(page);
3029                         kunmap_atomic(userpage);
3030
3031                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3032                                             &cached, GFP_NOFS);
3033                         unlock_extent_cached(tree, cur,
3034                                              cur + iosize - 1,
3035                                              &cached, GFP_NOFS);
3036                         cur = cur + iosize;
3037                         pg_offset += iosize;
3038                         continue;
3039                 }
3040                 /* the get_extent function already copied into the page */
3041                 if (test_range_bit(tree, cur, cur_end,
3042                                    EXTENT_UPTODATE, 1, NULL)) {
3043                         check_page_uptodate(tree, page);
3044                         unlock_extent(tree, cur, cur + iosize - 1);
3045                         cur = cur + iosize;
3046                         pg_offset += iosize;
3047                         continue;
3048                 }
3049                 /* we have an inline extent but it didn't get marked up
3050                  * to date.  Error out
3051                  */
3052                 if (block_start == EXTENT_MAP_INLINE) {
3053                         SetPageError(page);
3054                         unlock_extent(tree, cur, cur + iosize - 1);
3055                         cur = cur + iosize;
3056                         pg_offset += iosize;
3057                         continue;
3058                 }
3059
3060                 pnr -= page->index;
3061                 ret = submit_extent_page(rw, tree, NULL, page,
3062                                          sector, disk_io_size, pg_offset,
3063                                          bdev, bio, pnr,
3064                                          end_bio_extent_readpage, mirror_num,
3065                                          *bio_flags,
3066                                          this_bio_flag,
3067                                          force_bio_submit);
3068                 if (!ret) {
3069                         nr++;
3070                         *bio_flags = this_bio_flag;
3071                 } else {
3072                         SetPageError(page);
3073                         unlock_extent(tree, cur, cur + iosize - 1);
3074                 }
3075                 cur = cur + iosize;
3076                 pg_offset += iosize;
3077         }
3078 out:
3079         if (!nr) {
3080                 if (!PageError(page))
3081                         SetPageUptodate(page);
3082                 unlock_page(page);
3083         }
3084         return 0;
3085 }
3086
3087 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3088                                              struct page *pages[], int nr_pages,
3089                                              u64 start, u64 end,
3090                                              get_extent_t *get_extent,
3091                                              struct extent_map **em_cached,
3092                                              struct bio **bio, int mirror_num,
3093                                              unsigned long *bio_flags, int rw,
3094                                              u64 *prev_em_start)
3095 {
3096         struct inode *inode;
3097         struct btrfs_ordered_extent *ordered;
3098         int index;
3099
3100         inode = pages[0]->mapping->host;
3101         while (1) {
3102                 lock_extent(tree, start, end);
3103                 ordered = btrfs_lookup_ordered_range(inode, start,
3104                                                      end - start + 1);
3105                 if (!ordered)
3106                         break;
3107                 unlock_extent(tree, start, end);
3108                 btrfs_start_ordered_extent(inode, ordered, 1);
3109                 btrfs_put_ordered_extent(ordered);
3110         }
3111
3112         for (index = 0; index < nr_pages; index++) {
3113                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3114                               mirror_num, bio_flags, rw, prev_em_start);
3115                 page_cache_release(pages[index]);
3116         }
3117 }
3118
3119 static void __extent_readpages(struct extent_io_tree *tree,
3120                                struct page *pages[],
3121                                int nr_pages, get_extent_t *get_extent,
3122                                struct extent_map **em_cached,
3123                                struct bio **bio, int mirror_num,
3124                                unsigned long *bio_flags, int rw,
3125                                u64 *prev_em_start)
3126 {
3127         u64 start = 0;
3128         u64 end = 0;
3129         u64 page_start;
3130         int index;
3131         int first_index = 0;
3132
3133         for (index = 0; index < nr_pages; index++) {
3134                 page_start = page_offset(pages[index]);
3135                 if (!end) {
3136                         start = page_start;
3137                         end = start + PAGE_CACHE_SIZE - 1;
3138                         first_index = index;
3139                 } else if (end + 1 == page_start) {
3140                         end += PAGE_CACHE_SIZE;
3141                 } else {
3142                         __do_contiguous_readpages(tree, &pages[first_index],
3143                                                   index - first_index, start,
3144                                                   end, get_extent, em_cached,
3145                                                   bio, mirror_num, bio_flags,
3146                                                   rw, prev_em_start);
3147                         start = page_start;
3148                         end = start + PAGE_CACHE_SIZE - 1;
3149                         first_index = index;
3150                 }
3151         }
3152
3153         if (end)
3154                 __do_contiguous_readpages(tree, &pages[first_index],
3155                                           index - first_index, start,
3156                                           end, get_extent, em_cached, bio,
3157                                           mirror_num, bio_flags, rw,
3158                                           prev_em_start);
3159 }
3160
3161 static int __extent_read_full_page(struct extent_io_tree *tree,
3162                                    struct page *page,
3163                                    get_extent_t *get_extent,
3164                                    struct bio **bio, int mirror_num,
3165                                    unsigned long *bio_flags, int rw)
3166 {
3167         struct inode *inode = page->mapping->host;
3168         struct btrfs_ordered_extent *ordered;
3169         u64 start = page_offset(page);
3170         u64 end = start + PAGE_CACHE_SIZE - 1;
3171         int ret;
3172
3173         while (1) {
3174                 lock_extent(tree, start, end);
3175                 ordered = btrfs_lookup_ordered_range(inode, start,
3176                                                 PAGE_CACHE_SIZE);
3177                 if (!ordered)
3178                         break;
3179                 unlock_extent(tree, start, end);
3180                 btrfs_start_ordered_extent(inode, ordered, 1);
3181                 btrfs_put_ordered_extent(ordered);
3182         }
3183
3184         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3185                             bio_flags, rw, NULL);
3186         return ret;
3187 }
3188
3189 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3190                             get_extent_t *get_extent, int mirror_num)
3191 {
3192         struct bio *bio = NULL;
3193         unsigned long bio_flags = 0;
3194         int ret;
3195
3196         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3197                                       &bio_flags, READ);
3198         if (bio)
3199                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3200         return ret;
3201 }
3202
3203 static noinline void update_nr_written(struct page *page,
3204                                       struct writeback_control *wbc,
3205                                       unsigned long nr_written)
3206 {
3207         wbc->nr_to_write -= nr_written;
3208         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3209             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3210                 page->mapping->writeback_index = page->index + nr_written;
3211 }
3212
3213 /*
3214  * helper for __extent_writepage, doing all of the delayed allocation setup.
3215  *
3216  * This returns 1 if our fill_delalloc function did all the work required
3217  * to write the page (copy into inline extent).  In this case the IO has
3218  * been started and the page is already unlocked.
3219  *
3220  * This returns 0 if all went well (page still locked)
3221  * This returns < 0 if there were errors (page still locked)
3222  */
3223 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3224                               struct page *page, struct writeback_control *wbc,
3225                               struct extent_page_data *epd,
3226                               u64 delalloc_start,
3227                               unsigned long *nr_written)
3228 {
3229         struct extent_io_tree *tree = epd->tree;
3230         u64 page_end = delalloc_start + PAGE_CACHE_SIZE - 1;
3231         u64 nr_delalloc;
3232         u64 delalloc_to_write = 0;
3233         u64 delalloc_end = 0;
3234         int ret;
3235         int page_started = 0;
3236
3237         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3238                 return 0;
3239
3240         while (delalloc_end < page_end) {
3241                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3242                                                page,
3243                                                &delalloc_start,
3244                                                &delalloc_end,
3245                                                BTRFS_MAX_EXTENT_SIZE);
3246                 if (nr_delalloc == 0) {
3247                         delalloc_start = delalloc_end + 1;
3248                         continue;
3249                 }
3250                 ret = tree->ops->fill_delalloc(inode, page,
3251                                                delalloc_start,
3252                                                delalloc_end,
3253                                                &page_started,
3254                                                nr_written);
3255                 /* File system has been set read-only */
3256                 if (ret) {
3257                         SetPageError(page);
3258                         /* fill_delalloc should be return < 0 for error
3259                          * but just in case, we use > 0 here meaning the
3260                          * IO is started, so we don't want to return > 0
3261                          * unless things are going well.
3262                          */
3263                         ret = ret < 0 ? ret : -EIO;
3264                         goto done;
3265                 }
3266                 /*
3267                  * delalloc_end is already one less than the total
3268                  * length, so we don't subtract one from
3269                  * PAGE_CACHE_SIZE
3270                  */
3271                 delalloc_to_write += (delalloc_end - delalloc_start +
3272                                       PAGE_CACHE_SIZE) >>
3273                                       PAGE_CACHE_SHIFT;
3274                 delalloc_start = delalloc_end + 1;
3275         }
3276         if (wbc->nr_to_write < delalloc_to_write) {
3277                 int thresh = 8192;
3278
3279                 if (delalloc_to_write < thresh * 2)
3280                         thresh = delalloc_to_write;
3281                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3282                                          thresh);
3283         }
3284
3285         /* did the fill delalloc function already unlock and start
3286          * the IO?
3287          */
3288         if (page_started) {
3289                 /*
3290                  * we've unlocked the page, so we can't update
3291                  * the mapping's writeback index, just update
3292                  * nr_to_write.
3293                  */
3294                 wbc->nr_to_write -= *nr_written;
3295                 return 1;
3296         }
3297
3298         ret = 0;
3299
3300 done:
3301         return ret;
3302 }
3303
3304 /*
3305  * helper for __extent_writepage.  This calls the writepage start hooks,
3306  * and does the loop to map the page into extents and bios.
3307  *
3308  * We return 1 if the IO is started and the page is unlocked,
3309  * 0 if all went well (page still locked)
3310  * < 0 if there were errors (page still locked)
3311  */
3312 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3313                                  struct page *page,
3314                                  struct writeback_control *wbc,
3315                                  struct extent_page_data *epd,
3316                                  loff_t i_size,
3317                                  unsigned long nr_written,
3318                                  int write_flags, int *nr_ret)
3319 {
3320         struct extent_io_tree *tree = epd->tree;
3321         u64 start = page_offset(page);
3322         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3323         u64 end;
3324         u64 cur = start;
3325         u64 extent_offset;
3326         u64 block_start;
3327         u64 iosize;
3328         sector_t sector;
3329         struct extent_state *cached_state = NULL;
3330         struct extent_map *em;
3331         struct block_device *bdev;
3332         size_t pg_offset = 0;
3333         size_t blocksize;
3334         int ret = 0;
3335         int nr = 0;
3336         bool compressed;
3337
3338         if (tree->ops && tree->ops->writepage_start_hook) {
3339                 ret = tree->ops->writepage_start_hook(page, start,
3340                                                       page_end);
3341                 if (ret) {
3342                         /* Fixup worker will requeue */
3343                         if (ret == -EBUSY)
3344                                 wbc->pages_skipped++;
3345                         else
3346                                 redirty_page_for_writepage(wbc, page);
3347
3348                         update_nr_written(page, wbc, nr_written);
3349                         unlock_page(page);
3350                         ret = 1;
3351                         goto done_unlocked;
3352                 }
3353         }
3354
3355         /*
3356          * we don't want to touch the inode after unlocking the page,
3357          * so we update the mapping writeback index now
3358          */
3359         update_nr_written(page, wbc, nr_written + 1);
3360
3361         end = page_end;
3362         if (i_size <= start) {
3363                 if (tree->ops && tree->ops->writepage_end_io_hook)
3364                         tree->ops->writepage_end_io_hook(page, start,
3365                                                          page_end, NULL, 1);
3366                 goto done;
3367         }
3368
3369         blocksize = inode->i_sb->s_blocksize;
3370
3371         while (cur <= end) {
3372                 u64 em_end;
3373                 if (cur >= i_size) {
3374                         if (tree->ops && tree->ops->writepage_end_io_hook)
3375                                 tree->ops->writepage_end_io_hook(page, cur,
3376                                                          page_end, NULL, 1);
3377                         break;
3378                 }
3379                 em = epd->get_extent(inode, page, pg_offset, cur,
3380                                      end - cur + 1, 1);
3381                 if (IS_ERR_OR_NULL(em)) {
3382                         SetPageError(page);
3383                         ret = PTR_ERR_OR_ZERO(em);
3384                         break;
3385                 }
3386
3387                 extent_offset = cur - em->start;
3388                 em_end = extent_map_end(em);
3389                 BUG_ON(em_end <= cur);
3390                 BUG_ON(end < cur);
3391                 iosize = min(em_end - cur, end - cur + 1);
3392                 iosize = ALIGN(iosize, blocksize);
3393                 sector = (em->block_start + extent_offset) >> 9;
3394                 bdev = em->bdev;
3395                 block_start = em->block_start;
3396                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3397                 free_extent_map(em);
3398                 em = NULL;
3399
3400                 /*
3401                  * compressed and inline extents are written through other
3402                  * paths in the FS
3403                  */
3404                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3405                     block_start == EXTENT_MAP_INLINE) {
3406                         /*
3407                          * end_io notification does not happen here for
3408                          * compressed extents
3409                          */
3410                         if (!compressed && tree->ops &&
3411                             tree->ops->writepage_end_io_hook)
3412                                 tree->ops->writepage_end_io_hook(page, cur,
3413                                                          cur + iosize - 1,
3414                                                          NULL, 1);
3415                         else if (compressed) {
3416                                 /* we don't want to end_page_writeback on
3417                                  * a compressed extent.  this happens
3418                                  * elsewhere
3419                                  */
3420                                 nr++;
3421                         }
3422
3423                         cur += iosize;
3424                         pg_offset += iosize;
3425                         continue;
3426                 }
3427
3428                 if (tree->ops && tree->ops->writepage_io_hook) {
3429                         ret = tree->ops->writepage_io_hook(page, cur,
3430                                                 cur + iosize - 1);
3431                 } else {
3432                         ret = 0;
3433                 }
3434                 if (ret) {
3435                         SetPageError(page);
3436                 } else {
3437                         unsigned long max_nr = (i_size >> PAGE_CACHE_SHIFT) + 1;
3438
3439                         set_range_writeback(tree, cur, cur + iosize - 1);
3440                         if (!PageWriteback(page)) {
3441                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
3442                                            "page %lu not writeback, cur %llu end %llu",
3443                                        page->index, cur, end);
3444                         }
3445
3446                         ret = submit_extent_page(write_flags, tree, wbc, page,
3447                                                  sector, iosize, pg_offset,
3448                                                  bdev, &epd->bio, max_nr,
3449                                                  end_bio_extent_writepage,
3450                                                  0, 0, 0, false);
3451                         if (ret)
3452                                 SetPageError(page);
3453                 }
3454                 cur = cur + iosize;
3455                 pg_offset += iosize;
3456                 nr++;
3457         }
3458 done:
3459         *nr_ret = nr;
3460
3461 done_unlocked:
3462
3463         /* drop our reference on any cached states */
3464         free_extent_state(cached_state);
3465         return ret;
3466 }
3467
3468 /*
3469  * the writepage semantics are similar to regular writepage.  extent
3470  * records are inserted to lock ranges in the tree, and as dirty areas
3471  * are found, they are marked writeback.  Then the lock bits are removed
3472  * and the end_io handler clears the writeback ranges
3473  */
3474 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3475                               void *data)
3476 {
3477         struct inode *inode = page->mapping->host;
3478         struct extent_page_data *epd = data;
3479         u64 start = page_offset(page);
3480         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3481         int ret;
3482         int nr = 0;
3483         size_t pg_offset = 0;
3484         loff_t i_size = i_size_read(inode);
3485         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3486         int write_flags;
3487         unsigned long nr_written = 0;
3488
3489         if (wbc->sync_mode == WB_SYNC_ALL)
3490                 write_flags = WRITE_SYNC;
3491         else
3492                 write_flags = WRITE;
3493
3494         trace___extent_writepage(page, inode, wbc);
3495
3496         WARN_ON(!PageLocked(page));
3497
3498         ClearPageError(page);
3499
3500         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3501         if (page->index > end_index ||
3502            (page->index == end_index && !pg_offset)) {
3503                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3504                 unlock_page(page);
3505                 return 0;
3506         }
3507
3508         if (page->index == end_index) {
3509                 char *userpage;
3510
3511                 userpage = kmap_atomic(page);
3512                 memset(userpage + pg_offset, 0,
3513                        PAGE_CACHE_SIZE - pg_offset);
3514                 kunmap_atomic(userpage);
3515                 flush_dcache_page(page);
3516         }
3517
3518         pg_offset = 0;
3519
3520         set_page_extent_mapped(page);
3521
3522         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3523         if (ret == 1)
3524                 goto done_unlocked;
3525         if (ret)
3526                 goto done;
3527
3528         ret = __extent_writepage_io(inode, page, wbc, epd,
3529                                     i_size, nr_written, write_flags, &nr);
3530         if (ret == 1)
3531                 goto done_unlocked;
3532
3533 done:
3534         if (nr == 0) {
3535                 /* make sure the mapping tag for page dirty gets cleared */
3536                 set_page_writeback(page);
3537                 end_page_writeback(page);
3538         }
3539         if (PageError(page)) {
3540                 ret = ret < 0 ? ret : -EIO;
3541                 end_extent_writepage(page, ret, start, page_end);
3542         }
3543         unlock_page(page);
3544         return ret;
3545
3546 done_unlocked:
3547         return 0;
3548 }
3549
3550 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3551 {
3552         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3553                        TASK_UNINTERRUPTIBLE);
3554 }
3555
3556 static noinline_for_stack int
3557 lock_extent_buffer_for_io(struct extent_buffer *eb,
3558                           struct btrfs_fs_info *fs_info,
3559                           struct extent_page_data *epd)
3560 {
3561         unsigned long i, num_pages;
3562         int flush = 0;
3563         int ret = 0;
3564
3565         if (!btrfs_try_tree_write_lock(eb)) {
3566                 flush = 1;
3567                 flush_write_bio(epd);
3568                 btrfs_tree_lock(eb);
3569         }
3570
3571         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3572                 btrfs_tree_unlock(eb);
3573                 if (!epd->sync_io)
3574                         return 0;
3575                 if (!flush) {
3576                         flush_write_bio(epd);
3577                         flush = 1;
3578                 }
3579                 while (1) {
3580                         wait_on_extent_buffer_writeback(eb);
3581                         btrfs_tree_lock(eb);
3582                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3583                                 break;
3584                         btrfs_tree_unlock(eb);
3585                 }
3586         }
3587
3588         /*
3589          * We need to do this to prevent races in people who check if the eb is
3590          * under IO since we can end up having no IO bits set for a short period
3591          * of time.
3592          */
3593         spin_lock(&eb->refs_lock);
3594         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3595                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3596                 spin_unlock(&eb->refs_lock);
3597                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3598                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3599                                      -eb->len,
3600                                      fs_info->dirty_metadata_batch);
3601                 ret = 1;
3602         } else {
3603                 spin_unlock(&eb->refs_lock);
3604         }
3605
3606         btrfs_tree_unlock(eb);
3607
3608         if (!ret)
3609                 return ret;
3610
3611         num_pages = num_extent_pages(eb->start, eb->len);
3612         for (i = 0; i < num_pages; i++) {
3613                 struct page *p = eb->pages[i];
3614
3615                 if (!trylock_page(p)) {
3616                         if (!flush) {
3617                                 flush_write_bio(epd);
3618                                 flush = 1;
3619                         }
3620                         lock_page(p);
3621                 }
3622         }
3623
3624         return ret;
3625 }
3626
3627 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3628 {
3629         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3630         smp_mb__after_atomic();
3631         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3632 }
3633
3634 static void set_btree_ioerr(struct page *page)
3635 {
3636         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3637         struct btrfs_inode *btree_ino = BTRFS_I(eb->fs_info->btree_inode);
3638
3639         SetPageError(page);
3640         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3641                 return;
3642
3643         /*
3644          * If writeback for a btree extent that doesn't belong to a log tree
3645          * failed, increment the counter transaction->eb_write_errors.
3646          * We do this because while the transaction is running and before it's
3647          * committing (when we call filemap_fdata[write|wait]_range against
3648          * the btree inode), we might have
3649          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3650          * returns an error or an error happens during writeback, when we're
3651          * committing the transaction we wouldn't know about it, since the pages
3652          * can be no longer dirty nor marked anymore for writeback (if a
3653          * subsequent modification to the extent buffer didn't happen before the
3654          * transaction commit), which makes filemap_fdata[write|wait]_range not
3655          * able to find the pages tagged with SetPageError at transaction
3656          * commit time. So if this happens we must abort the transaction,
3657          * otherwise we commit a super block with btree roots that point to
3658          * btree nodes/leafs whose content on disk is invalid - either garbage
3659          * or the content of some node/leaf from a past generation that got
3660          * cowed or deleted and is no longer valid.
3661          *
3662          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3663          * not be enough - we need to distinguish between log tree extents vs
3664          * non-log tree extents, and the next filemap_fdatawait_range() call
3665          * will catch and clear such errors in the mapping - and that call might
3666          * be from a log sync and not from a transaction commit. Also, checking
3667          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3668          * not done and would not be reliable - the eb might have been released
3669          * from memory and reading it back again means that flag would not be
3670          * set (since it's a runtime flag, not persisted on disk).
3671          *
3672          * Using the flags below in the btree inode also makes us achieve the
3673          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3674          * writeback for all dirty pages and before filemap_fdatawait_range()
3675          * is called, the writeback for all dirty pages had already finished
3676          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3677          * filemap_fdatawait_range() would return success, as it could not know
3678          * that writeback errors happened (the pages were no longer tagged for
3679          * writeback).
3680          */
3681         switch (eb->log_index) {
3682         case -1:
3683                 set_bit(BTRFS_INODE_BTREE_ERR, &btree_ino->runtime_flags);
3684                 break;
3685         case 0:
3686                 set_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
3687                 break;
3688         case 1:
3689                 set_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
3690                 break;
3691         default:
3692                 BUG(); /* unexpected, logic error */
3693         }
3694 }
3695
3696 static void end_bio_extent_buffer_writepage(struct bio *bio)
3697 {
3698         struct bio_vec *bvec;
3699         struct extent_buffer *eb;
3700         int i, done;
3701
3702         bio_for_each_segment_all(bvec, bio, i) {
3703                 struct page *page = bvec->bv_page;
3704
3705                 eb = (struct extent_buffer *)page->private;
3706                 BUG_ON(!eb);
3707                 done = atomic_dec_and_test(&eb->io_pages);
3708
3709                 if (bio->bi_error ||
3710                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3711                         ClearPageUptodate(page);
3712                         set_btree_ioerr(page);
3713                 }
3714
3715                 end_page_writeback(page);
3716
3717                 if (!done)
3718                         continue;
3719
3720                 end_extent_buffer_writeback(eb);
3721         }
3722
3723         bio_put(bio);
3724 }
3725
3726 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3727                         struct btrfs_fs_info *fs_info,
3728                         struct writeback_control *wbc,
3729                         struct extent_page_data *epd)
3730 {
3731         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3732         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3733         u64 offset = eb->start;
3734         unsigned long i, num_pages;
3735         unsigned long bio_flags = 0;
3736         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3737         int ret = 0;
3738
3739         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3740         num_pages = num_extent_pages(eb->start, eb->len);
3741         atomic_set(&eb->io_pages, num_pages);
3742         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3743                 bio_flags = EXTENT_BIO_TREE_LOG;
3744
3745         for (i = 0; i < num_pages; i++) {
3746                 struct page *p = eb->pages[i];
3747
3748                 clear_page_dirty_for_io(p);
3749                 set_page_writeback(p);
3750                 ret = submit_extent_page(rw, tree, wbc, p, offset >> 9,
3751                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3752                                          -1, end_bio_extent_buffer_writepage,
3753                                          0, epd->bio_flags, bio_flags, false);
3754                 epd->bio_flags = bio_flags;
3755                 if (ret) {
3756                         set_btree_ioerr(p);
3757                         end_page_writeback(p);
3758                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3759                                 end_extent_buffer_writeback(eb);
3760                         ret = -EIO;
3761                         break;
3762                 }
3763                 offset += PAGE_CACHE_SIZE;
3764                 update_nr_written(p, wbc, 1);
3765                 unlock_page(p);
3766         }
3767
3768         if (unlikely(ret)) {
3769                 for (; i < num_pages; i++) {
3770                         struct page *p = eb->pages[i];
3771                         clear_page_dirty_for_io(p);
3772                         unlock_page(p);
3773                 }
3774         }
3775
3776         return ret;
3777 }
3778
3779 int btree_write_cache_pages(struct address_space *mapping,
3780                                    struct writeback_control *wbc)
3781 {
3782         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3783         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3784         struct extent_buffer *eb, *prev_eb = NULL;
3785         struct extent_page_data epd = {
3786                 .bio = NULL,
3787                 .tree = tree,
3788                 .extent_locked = 0,
3789                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3790                 .bio_flags = 0,
3791         };
3792         int ret = 0;
3793         int done = 0;
3794         int nr_to_write_done = 0;
3795         struct pagevec pvec;
3796         int nr_pages;
3797         pgoff_t index;
3798         pgoff_t end;            /* Inclusive */
3799         int scanned = 0;
3800         int tag;
3801
3802         pagevec_init(&pvec, 0);
3803         if (wbc->range_cyclic) {
3804                 index = mapping->writeback_index; /* Start from prev offset */
3805                 end = -1;
3806         } else {
3807                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3808                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3809                 scanned = 1;
3810         }
3811         if (wbc->sync_mode == WB_SYNC_ALL)
3812                 tag = PAGECACHE_TAG_TOWRITE;
3813         else
3814                 tag = PAGECACHE_TAG_DIRTY;
3815 retry:
3816         if (wbc->sync_mode == WB_SYNC_ALL)
3817                 tag_pages_for_writeback(mapping, index, end);
3818         while (!done && !nr_to_write_done && (index <= end) &&
3819                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3820                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3821                 unsigned i;
3822
3823                 scanned = 1;
3824                 for (i = 0; i < nr_pages; i++) {
3825                         struct page *page = pvec.pages[i];
3826
3827                         if (!PagePrivate(page))
3828                                 continue;
3829
3830                         if (!wbc->range_cyclic && page->index > end) {
3831                                 done = 1;
3832                                 break;
3833                         }
3834
3835                         spin_lock(&mapping->private_lock);
3836                         if (!PagePrivate(page)) {
3837                                 spin_unlock(&mapping->private_lock);
3838                                 continue;
3839                         }
3840
3841                         eb = (struct extent_buffer *)page->private;
3842
3843                         /*
3844                          * Shouldn't happen and normally this would be a BUG_ON
3845                          * but no sense in crashing the users box for something
3846                          * we can survive anyway.
3847                          */
3848                         if (WARN_ON(!eb)) {
3849                                 spin_unlock(&mapping->private_lock);
3850                                 continue;
3851                         }
3852
3853                         if (eb == prev_eb) {
3854                                 spin_unlock(&mapping->private_lock);
3855                                 continue;
3856                         }
3857
3858                         ret = atomic_inc_not_zero(&eb->refs);
3859                         spin_unlock(&mapping->private_lock);
3860                         if (!ret)
3861                                 continue;
3862
3863                         prev_eb = eb;
3864                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3865                         if (!ret) {
3866                                 free_extent_buffer(eb);
3867                                 continue;
3868                         }
3869
3870                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3871                         if (ret) {
3872                                 done = 1;
3873                                 free_extent_buffer(eb);
3874                                 break;
3875                         }
3876                         free_extent_buffer(eb);
3877
3878                         /*
3879                          * the filesystem may choose to bump up nr_to_write.
3880                          * We have to make sure to honor the new nr_to_write
3881                          * at any time
3882                          */
3883                         nr_to_write_done = wbc->nr_to_write <= 0;
3884                 }
3885                 pagevec_release(&pvec);
3886                 cond_resched();
3887         }
3888         if (!scanned && !done) {
3889                 /*
3890                  * We hit the last page and there is more work to be done: wrap
3891                  * back to the start of the file
3892                  */
3893                 scanned = 1;
3894                 index = 0;
3895                 goto retry;
3896         }
3897         flush_write_bio(&epd);
3898         return ret;
3899 }
3900
3901 /**
3902  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3903  * @mapping: address space structure to write
3904  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3905  * @writepage: function called for each page
3906  * @data: data passed to writepage function
3907  *
3908  * If a page is already under I/O, write_cache_pages() skips it, even
3909  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3910  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3911  * and msync() need to guarantee that all the data which was dirty at the time
3912  * the call was made get new I/O started against them.  If wbc->sync_mode is
3913  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3914  * existing IO to complete.
3915  */
3916 static int extent_write_cache_pages(struct extent_io_tree *tree,
3917                              struct address_space *mapping,
3918                              struct writeback_control *wbc,
3919                              writepage_t writepage, void *data,
3920                              void (*flush_fn)(void *))
3921 {
3922         struct inode *inode = mapping->host;
3923         int ret = 0;
3924         int done = 0;
3925         int err = 0;
3926         int nr_to_write_done = 0;
3927         struct pagevec pvec;
3928         int nr_pages;
3929         pgoff_t index;
3930         pgoff_t end;            /* Inclusive */
3931         int scanned = 0;
3932         int tag;
3933
3934         /*
3935          * We have to hold onto the inode so that ordered extents can do their
3936          * work when the IO finishes.  The alternative to this is failing to add
3937          * an ordered extent if the igrab() fails there and that is a huge pain
3938          * to deal with, so instead just hold onto the inode throughout the
3939          * writepages operation.  If it fails here we are freeing up the inode
3940          * anyway and we'd rather not waste our time writing out stuff that is
3941          * going to be truncated anyway.
3942          */
3943         if (!igrab(inode))
3944                 return 0;
3945
3946         pagevec_init(&pvec, 0);
3947         if (wbc->range_cyclic) {
3948                 index = mapping->writeback_index; /* Start from prev offset */
3949                 end = -1;
3950         } else {
3951                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3952                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3953                 scanned = 1;
3954         }
3955         if (wbc->sync_mode == WB_SYNC_ALL)
3956                 tag = PAGECACHE_TAG_TOWRITE;
3957         else
3958                 tag = PAGECACHE_TAG_DIRTY;
3959 retry:
3960         if (wbc->sync_mode == WB_SYNC_ALL)
3961                 tag_pages_for_writeback(mapping, index, end);
3962         while (!done && !nr_to_write_done && (index <= end) &&
3963                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3964                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3965                 unsigned i;
3966
3967                 scanned = 1;
3968                 for (i = 0; i < nr_pages; i++) {
3969                         struct page *page = pvec.pages[i];
3970
3971                         /*
3972                          * At this point we hold neither mapping->tree_lock nor
3973                          * lock on the page itself: the page may be truncated or
3974                          * invalidated (changing page->mapping to NULL), or even
3975                          * swizzled back from swapper_space to tmpfs file
3976                          * mapping
3977                          */
3978                         if (!trylock_page(page)) {
3979                                 flush_fn(data);
3980                                 lock_page(page);
3981                         }
3982
3983                         if (unlikely(page->mapping != mapping)) {
3984                                 unlock_page(page);
3985                                 continue;
3986                         }
3987
3988                         if (!wbc->range_cyclic && page->index > end) {
3989                                 done = 1;
3990                                 unlock_page(page);
3991                                 continue;
3992                         }
3993
3994                         if (wbc->sync_mode != WB_SYNC_NONE) {
3995                                 if (PageWriteback(page))
3996                                         flush_fn(data);
3997                                 wait_on_page_writeback(page);
3998                         }
3999
4000                         if (PageWriteback(page) ||
4001                             !clear_page_dirty_for_io(page)) {
4002                                 unlock_page(page);
4003                                 continue;
4004                         }
4005
4006                         ret = (*writepage)(page, wbc, data);
4007
4008                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4009                                 unlock_page(page);
4010                                 ret = 0;
4011                         }
4012                         if (!err && ret < 0)
4013                                 err = ret;
4014
4015                         /*
4016                          * the filesystem may choose to bump up nr_to_write.
4017                          * We have to make sure to honor the new nr_to_write
4018                          * at any time
4019                          */
4020                         nr_to_write_done = wbc->nr_to_write <= 0;
4021                 }
4022                 pagevec_release(&pvec);
4023                 cond_resched();
4024         }
4025         if (!scanned && !done && !err) {
4026                 /*
4027                  * We hit the last page and there is more work to be done: wrap
4028                  * back to the start of the file
4029                  */
4030                 scanned = 1;
4031                 index = 0;
4032                 goto retry;
4033         }
4034         btrfs_add_delayed_iput(inode);
4035         return err;
4036 }
4037
4038 static void flush_epd_write_bio(struct extent_page_data *epd)
4039 {
4040         if (epd->bio) {
4041                 int rw = WRITE;
4042                 int ret;
4043
4044                 if (epd->sync_io)
4045                         rw = WRITE_SYNC;
4046
4047                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
4048                 BUG_ON(ret < 0); /* -ENOMEM */
4049                 epd->bio = NULL;
4050         }
4051 }
4052
4053 static noinline void flush_write_bio(void *data)
4054 {
4055         struct extent_page_data *epd = data;
4056         flush_epd_write_bio(epd);
4057 }
4058
4059 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4060                           get_extent_t *get_extent,
4061                           struct writeback_control *wbc)
4062 {
4063         int ret;
4064         struct extent_page_data epd = {
4065                 .bio = NULL,
4066                 .tree = tree,
4067                 .get_extent = get_extent,
4068                 .extent_locked = 0,
4069                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4070                 .bio_flags = 0,
4071         };
4072
4073         ret = __extent_writepage(page, wbc, &epd);
4074
4075         flush_epd_write_bio(&epd);
4076         return ret;
4077 }
4078
4079 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4080                               u64 start, u64 end, get_extent_t *get_extent,
4081                               int mode)
4082 {
4083         int ret = 0;
4084         struct address_space *mapping = inode->i_mapping;
4085         struct page *page;
4086         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
4087                 PAGE_CACHE_SHIFT;
4088
4089         struct extent_page_data epd = {
4090                 .bio = NULL,
4091                 .tree = tree,
4092                 .get_extent = get_extent,
4093                 .extent_locked = 1,
4094                 .sync_io = mode == WB_SYNC_ALL,
4095                 .bio_flags = 0,
4096         };
4097         struct writeback_control wbc_writepages = {
4098                 .sync_mode      = mode,
4099                 .nr_to_write    = nr_pages * 2,
4100                 .range_start    = start,
4101                 .range_end      = end + 1,
4102         };
4103
4104         while (start <= end) {
4105                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
4106                 if (clear_page_dirty_for_io(page))
4107                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4108                 else {
4109                         if (tree->ops && tree->ops->writepage_end_io_hook)
4110                                 tree->ops->writepage_end_io_hook(page, start,
4111                                                  start + PAGE_CACHE_SIZE - 1,
4112                                                  NULL, 1);
4113                         unlock_page(page);
4114                 }
4115                 page_cache_release(page);
4116                 start += PAGE_CACHE_SIZE;
4117         }
4118
4119         flush_epd_write_bio(&epd);
4120         return ret;
4121 }
4122
4123 int extent_writepages(struct extent_io_tree *tree,
4124                       struct address_space *mapping,
4125                       get_extent_t *get_extent,
4126                       struct writeback_control *wbc)
4127 {
4128         int ret = 0;
4129         struct extent_page_data epd = {
4130                 .bio = NULL,
4131                 .tree = tree,
4132                 .get_extent = get_extent,
4133                 .extent_locked = 0,
4134                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4135                 .bio_flags = 0,
4136         };
4137
4138         ret = extent_write_cache_pages(tree, mapping, wbc,
4139                                        __extent_writepage, &epd,
4140                                        flush_write_bio);
4141         flush_epd_write_bio(&epd);
4142         return ret;
4143 }
4144
4145 int extent_readpages(struct extent_io_tree *tree,
4146                      struct address_space *mapping,
4147                      struct list_head *pages, unsigned nr_pages,
4148                      get_extent_t get_extent)
4149 {
4150         struct bio *bio = NULL;
4151         unsigned page_idx;
4152         unsigned long bio_flags = 0;
4153         struct page *pagepool[16];
4154         struct page *page;
4155         struct extent_map *em_cached = NULL;
4156         int nr = 0;
4157         u64 prev_em_start = (u64)-1;
4158
4159         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4160                 page = list_entry(pages->prev, struct page, lru);
4161
4162                 prefetchw(&page->flags);
4163                 list_del(&page->lru);
4164                 if (add_to_page_cache_lru(page, mapping,
4165                                         page->index, GFP_NOFS)) {
4166                         page_cache_release(page);
4167                         continue;
4168                 }
4169
4170                 pagepool[nr++] = page;
4171                 if (nr < ARRAY_SIZE(pagepool))
4172                         continue;
4173                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4174                                    &bio, 0, &bio_flags, READ, &prev_em_start);
4175                 nr = 0;
4176         }
4177         if (nr)
4178                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4179                                    &bio, 0, &bio_flags, READ, &prev_em_start);
4180
4181         if (em_cached)
4182                 free_extent_map(em_cached);
4183
4184         BUG_ON(!list_empty(pages));
4185         if (bio)
4186                 return submit_one_bio(READ, bio, 0, bio_flags);
4187         return 0;
4188 }
4189
4190 /*
4191  * basic invalidatepage code, this waits on any locked or writeback
4192  * ranges corresponding to the page, and then deletes any extent state
4193  * records from the tree
4194  */
4195 int extent_invalidatepage(struct extent_io_tree *tree,
4196                           struct page *page, unsigned long offset)
4197 {
4198         struct extent_state *cached_state = NULL;
4199         u64 start = page_offset(page);
4200         u64 end = start + PAGE_CACHE_SIZE - 1;
4201         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4202
4203         start += ALIGN(offset, blocksize);
4204         if (start > end)
4205                 return 0;
4206
4207         lock_extent_bits(tree, start, end, &cached_state);
4208         wait_on_page_writeback(page);
4209         clear_extent_bit(tree, start, end,
4210                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4211                          EXTENT_DO_ACCOUNTING,
4212                          1, 1, &cached_state, GFP_NOFS);
4213         return 0;
4214 }
4215
4216 /*
4217  * a helper for releasepage, this tests for areas of the page that
4218  * are locked or under IO and drops the related state bits if it is safe
4219  * to drop the page.
4220  */
4221 static int try_release_extent_state(struct extent_map_tree *map,
4222                                     struct extent_io_tree *tree,
4223                                     struct page *page, gfp_t mask)
4224 {
4225         u64 start = page_offset(page);
4226         u64 end = start + PAGE_CACHE_SIZE - 1;
4227         int ret = 1;
4228
4229         if (test_range_bit(tree, start, end,
4230                            EXTENT_IOBITS, 0, NULL))
4231                 ret = 0;
4232         else {
4233                 if ((mask & GFP_NOFS) == GFP_NOFS)
4234                         mask = GFP_NOFS;
4235                 /*
4236                  * at this point we can safely clear everything except the
4237                  * locked bit and the nodatasum bit
4238                  */
4239                 ret = clear_extent_bit(tree, start, end,
4240                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4241                                  0, 0, NULL, mask);
4242
4243                 /* if clear_extent_bit failed for enomem reasons,
4244                  * we can't allow the release to continue.
4245                  */
4246                 if (ret < 0)
4247                         ret = 0;
4248                 else
4249                         ret = 1;
4250         }
4251         return ret;
4252 }
4253
4254 /*
4255  * a helper for releasepage.  As long as there are no locked extents
4256  * in the range corresponding to the page, both state records and extent
4257  * map records are removed
4258  */
4259 int try_release_extent_mapping(struct extent_map_tree *map,
4260                                struct extent_io_tree *tree, struct page *page,
4261                                gfp_t mask)
4262 {
4263         struct extent_map *em;
4264         u64 start = page_offset(page);
4265         u64 end = start + PAGE_CACHE_SIZE - 1;
4266
4267         if (gfpflags_allow_blocking(mask) &&
4268             page->mapping->host->i_size > SZ_16M) {
4269                 u64 len;
4270                 while (start <= end) {
4271                         len = end - start + 1;
4272                         write_lock(&map->lock);
4273                         em = lookup_extent_mapping(map, start, len);
4274                         if (!em) {
4275                                 write_unlock(&map->lock);
4276                                 break;
4277                         }
4278                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4279                             em->start != start) {
4280                                 write_unlock(&map->lock);
4281                                 free_extent_map(em);
4282                                 break;
4283                         }
4284                         if (!test_range_bit(tree, em->start,
4285                                             extent_map_end(em) - 1,
4286                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4287                                             0, NULL)) {
4288                                 remove_extent_mapping(map, em);
4289                                 /* once for the rb tree */
4290                                 free_extent_map(em);
4291                         }
4292                         start = extent_map_end(em);
4293                         write_unlock(&map->lock);
4294
4295                         /* once for us */
4296                         free_extent_map(em);
4297                 }
4298         }
4299         return try_release_extent_state(map, tree, page, mask);
4300 }
4301
4302 /*
4303  * helper function for fiemap, which doesn't want to see any holes.
4304  * This maps until we find something past 'last'
4305  */
4306 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4307                                                 u64 offset,
4308                                                 u64 last,
4309                                                 get_extent_t *get_extent)
4310 {
4311         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4312         struct extent_map *em;
4313         u64 len;
4314
4315         if (offset >= last)
4316                 return NULL;
4317
4318         while (1) {
4319                 len = last - offset;
4320                 if (len == 0)
4321                         break;
4322                 len = ALIGN(len, sectorsize);
4323                 em = get_extent(inode, NULL, 0, offset, len, 0);
4324                 if (IS_ERR_OR_NULL(em))
4325                         return em;
4326
4327                 /* if this isn't a hole return it */
4328                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4329                     em->block_start != EXTENT_MAP_HOLE) {
4330                         return em;
4331                 }
4332
4333                 /* this is a hole, advance to the next extent */
4334                 offset = extent_map_end(em);
4335                 free_extent_map(em);
4336                 if (offset >= last)
4337                         break;
4338         }
4339         return NULL;
4340 }
4341
4342 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4343                 __u64 start, __u64 len, get_extent_t *get_extent)
4344 {
4345         int ret = 0;
4346         u64 off = start;
4347         u64 max = start + len;
4348         u32 flags = 0;
4349         u32 found_type;
4350         u64 last;
4351         u64 last_for_get_extent = 0;
4352         u64 disko = 0;
4353         u64 isize = i_size_read(inode);
4354         struct btrfs_key found_key;
4355         struct extent_map *em = NULL;
4356         struct extent_state *cached_state = NULL;
4357         struct btrfs_path *path;
4358         struct btrfs_root *root = BTRFS_I(inode)->root;
4359         int end = 0;
4360         u64 em_start = 0;
4361         u64 em_len = 0;
4362         u64 em_end = 0;
4363
4364         if (len == 0)
4365                 return -EINVAL;
4366
4367         path = btrfs_alloc_path();
4368         if (!path)
4369                 return -ENOMEM;
4370         path->leave_spinning = 1;
4371
4372         start = round_down(start, BTRFS_I(inode)->root->sectorsize);
4373         len = round_up(max, BTRFS_I(inode)->root->sectorsize) - start;
4374
4375         /*
4376          * lookup the last file extent.  We're not using i_size here
4377          * because there might be preallocation past i_size
4378          */
4379         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode), -1,
4380                                        0);
4381         if (ret < 0) {
4382                 btrfs_free_path(path);
4383                 return ret;
4384         }
4385         WARN_ON(!ret);
4386         path->slots[0]--;
4387         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4388         found_type = found_key.type;
4389
4390         /* No extents, but there might be delalloc bits */
4391         if (found_key.objectid != btrfs_ino(inode) ||
4392             found_type != BTRFS_EXTENT_DATA_KEY) {
4393                 /* have to trust i_size as the end */
4394                 last = (u64)-1;
4395                 last_for_get_extent = isize;
4396         } else {
4397                 /*
4398                  * remember the start of the last extent.  There are a
4399                  * bunch of different factors that go into the length of the
4400                  * extent, so its much less complex to remember where it started
4401                  */
4402                 last = found_key.offset;
4403                 last_for_get_extent = last + 1;
4404         }
4405         btrfs_release_path(path);
4406
4407         /*
4408          * we might have some extents allocated but more delalloc past those
4409          * extents.  so, we trust isize unless the start of the last extent is
4410          * beyond isize
4411          */
4412         if (last < isize) {
4413                 last = (u64)-1;
4414                 last_for_get_extent = isize;
4415         }
4416
4417         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4418                          &cached_state);
4419
4420         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4421                                    get_extent);
4422         if (!em)
4423                 goto out;
4424         if (IS_ERR(em)) {
4425                 ret = PTR_ERR(em);
4426                 goto out;
4427         }
4428
4429         while (!end) {
4430                 u64 offset_in_extent = 0;
4431
4432                 /* break if the extent we found is outside the range */
4433                 if (em->start >= max || extent_map_end(em) < off)
4434                         break;
4435
4436                 /*
4437                  * get_extent may return an extent that starts before our
4438                  * requested range.  We have to make sure the ranges
4439                  * we return to fiemap always move forward and don't
4440                  * overlap, so adjust the offsets here
4441                  */
4442                 em_start = max(em->start, off);
4443
4444                 /*
4445                  * record the offset from the start of the extent
4446                  * for adjusting the disk offset below.  Only do this if the
4447                  * extent isn't compressed since our in ram offset may be past
4448                  * what we have actually allocated on disk.
4449                  */
4450                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4451                         offset_in_extent = em_start - em->start;
4452                 em_end = extent_map_end(em);
4453                 em_len = em_end - em_start;
4454                 disko = 0;
4455                 flags = 0;
4456
4457                 /*
4458                  * bump off for our next call to get_extent
4459                  */
4460                 off = extent_map_end(em);
4461                 if (off >= max)
4462                         end = 1;
4463
4464                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4465                         end = 1;
4466                         flags |= FIEMAP_EXTENT_LAST;
4467                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4468                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4469                                   FIEMAP_EXTENT_NOT_ALIGNED);
4470                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4471                         flags |= (FIEMAP_EXTENT_DELALLOC |
4472                                   FIEMAP_EXTENT_UNKNOWN);
4473                 } else if (fieinfo->fi_extents_max) {
4474                         u64 bytenr = em->block_start -
4475                                 (em->start - em->orig_start);
4476
4477                         disko = em->block_start + offset_in_extent;
4478
4479                         /*
4480                          * As btrfs supports shared space, this information
4481                          * can be exported to userspace tools via
4482                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4483                          * then we're just getting a count and we can skip the
4484                          * lookup stuff.
4485                          */
4486                         ret = btrfs_check_shared(NULL, root->fs_info,
4487                                                  root->objectid,
4488                                                  btrfs_ino(inode), bytenr);
4489                         if (ret < 0)
4490                                 goto out_free;
4491                         if (ret)
4492                                 flags |= FIEMAP_EXTENT_SHARED;
4493                         ret = 0;
4494                 }
4495                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4496                         flags |= FIEMAP_EXTENT_ENCODED;
4497                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4498                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4499
4500                 free_extent_map(em);
4501                 em = NULL;
4502                 if ((em_start >= last) || em_len == (u64)-1 ||
4503                    (last == (u64)-1 && isize <= em_end)) {
4504                         flags |= FIEMAP_EXTENT_LAST;
4505                         end = 1;
4506                 }
4507
4508                 /* now scan forward to see if this is really the last extent. */
4509                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4510                                            get_extent);
4511                 if (IS_ERR(em)) {
4512                         ret = PTR_ERR(em);
4513                         goto out;
4514                 }
4515                 if (!em) {
4516                         flags |= FIEMAP_EXTENT_LAST;
4517                         end = 1;
4518                 }
4519                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4520                                               em_len, flags);
4521                 if (ret) {
4522                         if (ret == 1)
4523                                 ret = 0;
4524                         goto out_free;
4525                 }
4526         }
4527 out_free:
4528         free_extent_map(em);
4529 out:
4530         btrfs_free_path(path);
4531         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4532                              &cached_state, GFP_NOFS);
4533         return ret;
4534 }
4535
4536 static void __free_extent_buffer(struct extent_buffer *eb)
4537 {
4538         btrfs_leak_debug_del(&eb->leak_list);
4539         kmem_cache_free(extent_buffer_cache, eb);
4540 }
4541
4542 int extent_buffer_under_io(struct extent_buffer *eb)
4543 {
4544         return (atomic_read(&eb->io_pages) ||
4545                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4546                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4547 }
4548
4549 /*
4550  * Helper for releasing extent buffer page.
4551  */
4552 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4553 {
4554         unsigned long index;
4555         struct page *page;
4556         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4557
4558         BUG_ON(extent_buffer_under_io(eb));
4559
4560         index = num_extent_pages(eb->start, eb->len);
4561         if (index == 0)
4562                 return;
4563
4564         do {
4565                 index--;
4566                 page = eb->pages[index];
4567                 if (!page)
4568                         continue;
4569                 if (mapped)
4570                         spin_lock(&page->mapping->private_lock);
4571                 /*
4572                  * We do this since we'll remove the pages after we've
4573                  * removed the eb from the radix tree, so we could race
4574                  * and have this page now attached to the new eb.  So
4575                  * only clear page_private if it's still connected to
4576                  * this eb.
4577                  */
4578                 if (PagePrivate(page) &&
4579                     page->private == (unsigned long)eb) {
4580                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4581                         BUG_ON(PageDirty(page));
4582                         BUG_ON(PageWriteback(page));
4583                         /*
4584                          * We need to make sure we haven't be attached
4585                          * to a new eb.
4586                          */
4587                         ClearPagePrivate(page);
4588                         set_page_private(page, 0);
4589                         /* One for the page private */
4590                         page_cache_release(page);
4591                 }
4592
4593                 if (mapped)
4594                         spin_unlock(&page->mapping->private_lock);
4595
4596                 /* One for when we alloced the page */
4597                 page_cache_release(page);
4598         } while (index != 0);
4599 }
4600
4601 /*
4602  * Helper for releasing the extent buffer.
4603  */
4604 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4605 {
4606         btrfs_release_extent_buffer_page(eb);
4607         __free_extent_buffer(eb);
4608 }
4609
4610 static struct extent_buffer *
4611 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4612                       unsigned long len)
4613 {
4614         struct extent_buffer *eb = NULL;
4615
4616         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4617         eb->start = start;
4618         eb->len = len;
4619         eb->fs_info = fs_info;
4620         eb->bflags = 0;
4621         rwlock_init(&eb->lock);
4622         atomic_set(&eb->write_locks, 0);
4623         atomic_set(&eb->read_locks, 0);
4624         atomic_set(&eb->blocking_readers, 0);
4625         atomic_set(&eb->blocking_writers, 0);
4626         atomic_set(&eb->spinning_readers, 0);
4627         atomic_set(&eb->spinning_writers, 0);
4628         eb->lock_nested = 0;
4629         init_waitqueue_head(&eb->write_lock_wq);
4630         init_waitqueue_head(&eb->read_lock_wq);
4631
4632         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4633
4634         spin_lock_init(&eb->refs_lock);
4635         atomic_set(&eb->refs, 1);
4636         atomic_set(&eb->io_pages, 0);
4637
4638         /*
4639          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4640          */
4641         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4642                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4643         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4644
4645         return eb;
4646 }
4647
4648 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4649 {
4650         unsigned long i;
4651         struct page *p;
4652         struct extent_buffer *new;
4653         unsigned long num_pages = num_extent_pages(src->start, src->len);
4654
4655         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4656         if (new == NULL)
4657                 return NULL;
4658
4659         for (i = 0; i < num_pages; i++) {
4660                 p = alloc_page(GFP_NOFS);
4661                 if (!p) {
4662                         btrfs_release_extent_buffer(new);
4663                         return NULL;
4664                 }
4665                 attach_extent_buffer_page(new, p);
4666                 WARN_ON(PageDirty(p));
4667                 SetPageUptodate(p);
4668                 new->pages[i] = p;
4669         }
4670
4671         copy_extent_buffer(new, src, 0, 0, src->len);
4672         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4673         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4674
4675         return new;
4676 }
4677
4678 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4679                                                   u64 start, unsigned long len)
4680 {
4681         struct extent_buffer *eb;
4682         unsigned long num_pages;
4683         unsigned long i;
4684
4685         num_pages = num_extent_pages(start, len);
4686
4687         eb = __alloc_extent_buffer(fs_info, start, len);
4688         if (!eb)
4689                 return NULL;
4690
4691         for (i = 0; i < num_pages; i++) {
4692                 eb->pages[i] = alloc_page(GFP_NOFS);
4693                 if (!eb->pages[i])
4694                         goto err;
4695         }
4696         set_extent_buffer_uptodate(eb);
4697         btrfs_set_header_nritems(eb, 0);
4698         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4699
4700         return eb;
4701 err:
4702         for (; i > 0; i--)
4703                 __free_page(eb->pages[i - 1]);
4704         __free_extent_buffer(eb);
4705         return NULL;
4706 }
4707
4708 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4709                                                 u64 start)
4710 {
4711         unsigned long len;
4712
4713         if (!fs_info) {
4714                 /*
4715                  * Called only from tests that don't always have a fs_info
4716                  * available, but we know that nodesize is 4096
4717                  */
4718                 len = 4096;
4719         } else {
4720                 len = fs_info->tree_root->nodesize;
4721         }
4722
4723         return __alloc_dummy_extent_buffer(fs_info, start, len);
4724 }
4725
4726 static void check_buffer_tree_ref(struct extent_buffer *eb)
4727 {
4728         int refs;
4729         /* the ref bit is tricky.  We have to make sure it is set
4730          * if we have the buffer dirty.   Otherwise the
4731          * code to free a buffer can end up dropping a dirty
4732          * page
4733          *
4734          * Once the ref bit is set, it won't go away while the
4735          * buffer is dirty or in writeback, and it also won't
4736          * go away while we have the reference count on the
4737          * eb bumped.
4738          *
4739          * We can't just set the ref bit without bumping the
4740          * ref on the eb because free_extent_buffer might
4741          * see the ref bit and try to clear it.  If this happens
4742          * free_extent_buffer might end up dropping our original
4743          * ref by mistake and freeing the page before we are able
4744          * to add one more ref.
4745          *
4746          * So bump the ref count first, then set the bit.  If someone
4747          * beat us to it, drop the ref we added.
4748          */
4749         refs = atomic_read(&eb->refs);
4750         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4751                 return;
4752
4753         spin_lock(&eb->refs_lock);
4754         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4755                 atomic_inc(&eb->refs);
4756         spin_unlock(&eb->refs_lock);
4757 }
4758
4759 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4760                 struct page *accessed)
4761 {
4762         unsigned long num_pages, i;
4763
4764         check_buffer_tree_ref(eb);
4765
4766         num_pages = num_extent_pages(eb->start, eb->len);
4767         for (i = 0; i < num_pages; i++) {
4768                 struct page *p = eb->pages[i];
4769
4770                 if (p != accessed)
4771                         mark_page_accessed(p);
4772         }
4773 }
4774
4775 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4776                                          u64 start)
4777 {
4778         struct extent_buffer *eb;
4779
4780         rcu_read_lock();
4781         eb = radix_tree_lookup(&fs_info->buffer_radix,
4782                                start >> PAGE_CACHE_SHIFT);
4783         if (eb && atomic_inc_not_zero(&eb->refs)) {
4784                 rcu_read_unlock();
4785                 /*
4786                  * Lock our eb's refs_lock to avoid races with
4787                  * free_extent_buffer. When we get our eb it might be flagged
4788                  * with EXTENT_BUFFER_STALE and another task running
4789                  * free_extent_buffer might have seen that flag set,
4790                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4791                  * writeback flags not set) and it's still in the tree (flag
4792                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4793                  * of decrementing the extent buffer's reference count twice.
4794                  * So here we could race and increment the eb's reference count,
4795                  * clear its stale flag, mark it as dirty and drop our reference
4796                  * before the other task finishes executing free_extent_buffer,
4797                  * which would later result in an attempt to free an extent
4798                  * buffer that is dirty.
4799                  */
4800                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4801                         spin_lock(&eb->refs_lock);
4802                         spin_unlock(&eb->refs_lock);
4803                 }
4804                 mark_extent_buffer_accessed(eb, NULL);
4805                 return eb;
4806         }
4807         rcu_read_unlock();
4808
4809         return NULL;
4810 }
4811
4812 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4813 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4814                                                u64 start)
4815 {
4816         struct extent_buffer *eb, *exists = NULL;
4817         int ret;
4818
4819         eb = find_extent_buffer(fs_info, start);
4820         if (eb)
4821                 return eb;
4822         eb = alloc_dummy_extent_buffer(fs_info, start);
4823         if (!eb)
4824                 return NULL;
4825         eb->fs_info = fs_info;
4826 again:
4827         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4828         if (ret)
4829                 goto free_eb;
4830         spin_lock(&fs_info->buffer_lock);
4831         ret = radix_tree_insert(&fs_info->buffer_radix,
4832                                 start >> PAGE_CACHE_SHIFT, eb);
4833         spin_unlock(&fs_info->buffer_lock);
4834         radix_tree_preload_end();
4835         if (ret == -EEXIST) {
4836                 exists = find_extent_buffer(fs_info, start);
4837                 if (exists)
4838                         goto free_eb;
4839                 else
4840                         goto again;
4841         }
4842         check_buffer_tree_ref(eb);
4843         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4844
4845         /*
4846          * We will free dummy extent buffer's if they come into
4847          * free_extent_buffer with a ref count of 2, but if we are using this we
4848          * want the buffers to stay in memory until we're done with them, so
4849          * bump the ref count again.
4850          */
4851         atomic_inc(&eb->refs);
4852         return eb;
4853 free_eb:
4854         btrfs_release_extent_buffer(eb);
4855         return exists;
4856 }
4857 #endif
4858
4859 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4860                                           u64 start)
4861 {
4862         unsigned long len = fs_info->tree_root->nodesize;
4863         unsigned long num_pages = num_extent_pages(start, len);
4864         unsigned long i;
4865         unsigned long index = start >> PAGE_CACHE_SHIFT;
4866         struct extent_buffer *eb;
4867         struct extent_buffer *exists = NULL;
4868         struct page *p;
4869         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4870         int uptodate = 1;
4871         int ret;
4872
4873         eb = find_extent_buffer(fs_info, start);
4874         if (eb)
4875                 return eb;
4876
4877         eb = __alloc_extent_buffer(fs_info, start, len);
4878         if (!eb)
4879                 return NULL;
4880
4881         for (i = 0; i < num_pages; i++, index++) {
4882                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4883                 if (!p)
4884                         goto free_eb;
4885
4886                 spin_lock(&mapping->private_lock);
4887                 if (PagePrivate(p)) {
4888                         /*
4889                          * We could have already allocated an eb for this page
4890                          * and attached one so lets see if we can get a ref on
4891                          * the existing eb, and if we can we know it's good and
4892                          * we can just return that one, else we know we can just
4893                          * overwrite page->private.
4894                          */
4895                         exists = (struct extent_buffer *)p->private;
4896                         if (atomic_inc_not_zero(&exists->refs)) {
4897                                 spin_unlock(&mapping->private_lock);
4898                                 unlock_page(p);
4899                                 page_cache_release(p);
4900                                 mark_extent_buffer_accessed(exists, p);
4901                                 goto free_eb;
4902                         }
4903                         exists = NULL;
4904
4905                         /*
4906                          * Do this so attach doesn't complain and we need to
4907                          * drop the ref the old guy had.
4908                          */
4909                         ClearPagePrivate(p);
4910                         WARN_ON(PageDirty(p));
4911                         page_cache_release(p);
4912                 }
4913                 attach_extent_buffer_page(eb, p);
4914                 spin_unlock(&mapping->private_lock);
4915                 WARN_ON(PageDirty(p));
4916                 eb->pages[i] = p;
4917                 if (!PageUptodate(p))
4918                         uptodate = 0;
4919
4920                 /*
4921                  * see below about how we avoid a nasty race with release page
4922                  * and why we unlock later
4923                  */
4924         }
4925         if (uptodate)
4926                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4927 again:
4928         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4929         if (ret)
4930                 goto free_eb;
4931
4932         spin_lock(&fs_info->buffer_lock);
4933         ret = radix_tree_insert(&fs_info->buffer_radix,
4934                                 start >> PAGE_CACHE_SHIFT, eb);
4935         spin_unlock(&fs_info->buffer_lock);
4936         radix_tree_preload_end();
4937         if (ret == -EEXIST) {
4938                 exists = find_extent_buffer(fs_info, start);
4939                 if (exists)
4940                         goto free_eb;
4941                 else
4942                         goto again;
4943         }
4944         /* add one reference for the tree */
4945         check_buffer_tree_ref(eb);
4946         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4947
4948         /*
4949          * there is a race where release page may have
4950          * tried to find this extent buffer in the radix
4951          * but failed.  It will tell the VM it is safe to
4952          * reclaim the, and it will clear the page private bit.
4953          * We must make sure to set the page private bit properly
4954          * after the extent buffer is in the radix tree so
4955          * it doesn't get lost
4956          */
4957         SetPageChecked(eb->pages[0]);
4958         for (i = 1; i < num_pages; i++) {
4959                 p = eb->pages[i];
4960                 ClearPageChecked(p);
4961                 unlock_page(p);
4962         }
4963         unlock_page(eb->pages[0]);
4964         return eb;
4965
4966 free_eb:
4967         WARN_ON(!atomic_dec_and_test(&eb->refs));
4968         for (i = 0; i < num_pages; i++) {
4969                 if (eb->pages[i])
4970                         unlock_page(eb->pages[i]);
4971         }
4972
4973         btrfs_release_extent_buffer(eb);
4974         return exists;
4975 }
4976
4977 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4978 {
4979         struct extent_buffer *eb =
4980                         container_of(head, struct extent_buffer, rcu_head);
4981
4982         __free_extent_buffer(eb);
4983 }
4984
4985 /* Expects to have eb->eb_lock already held */
4986 static int release_extent_buffer(struct extent_buffer *eb)
4987 {
4988         WARN_ON(atomic_read(&eb->refs) == 0);
4989         if (atomic_dec_and_test(&eb->refs)) {
4990                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
4991                         struct btrfs_fs_info *fs_info = eb->fs_info;
4992
4993                         spin_unlock(&eb->refs_lock);
4994
4995                         spin_lock(&fs_info->buffer_lock);
4996                         radix_tree_delete(&fs_info->buffer_radix,
4997                                           eb->start >> PAGE_CACHE_SHIFT);
4998                         spin_unlock(&fs_info->buffer_lock);
4999                 } else {
5000                         spin_unlock(&eb->refs_lock);
5001                 }
5002
5003                 /* Should be safe to release our pages at this point */
5004                 btrfs_release_extent_buffer_page(eb);
5005 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5006                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5007                         __free_extent_buffer(eb);
5008                         return 1;
5009                 }
5010 #endif
5011                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5012                 return 1;
5013         }
5014         spin_unlock(&eb->refs_lock);
5015
5016         return 0;
5017 }
5018
5019 void free_extent_buffer(struct extent_buffer *eb)
5020 {
5021         int refs;
5022         int old;
5023         if (!eb)
5024                 return;
5025
5026         while (1) {
5027                 refs = atomic_read(&eb->refs);
5028                 if (refs <= 3)
5029                         break;
5030                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5031                 if (old == refs)
5032                         return;
5033         }
5034
5035         spin_lock(&eb->refs_lock);
5036         if (atomic_read(&eb->refs) == 2 &&
5037             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5038                 atomic_dec(&eb->refs);
5039
5040         if (atomic_read(&eb->refs) == 2 &&
5041             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5042             !extent_buffer_under_io(eb) &&
5043             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5044                 atomic_dec(&eb->refs);
5045
5046         /*
5047          * I know this is terrible, but it's temporary until we stop tracking
5048          * the uptodate bits and such for the extent buffers.
5049          */
5050         release_extent_buffer(eb);
5051 }
5052
5053 void free_extent_buffer_stale(struct extent_buffer *eb)
5054 {
5055         if (!eb)
5056                 return;
5057
5058         spin_lock(&eb->refs_lock);
5059         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5060
5061         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5062             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5063                 atomic_dec(&eb->refs);
5064         release_extent_buffer(eb);
5065 }
5066
5067 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5068 {
5069         unsigned long i;
5070         unsigned long num_pages;
5071         struct page *page;
5072
5073         num_pages = num_extent_pages(eb->start, eb->len);
5074
5075         for (i = 0; i < num_pages; i++) {
5076                 page = eb->pages[i];
5077                 if (!PageDirty(page))
5078                         continue;
5079
5080                 lock_page(page);
5081                 WARN_ON(!PagePrivate(page));
5082
5083                 clear_page_dirty_for_io(page);
5084                 spin_lock_irq(&page->mapping->tree_lock);
5085                 if (!PageDirty(page)) {
5086                         radix_tree_tag_clear(&page->mapping->page_tree,
5087                                                 page_index(page),
5088                                                 PAGECACHE_TAG_DIRTY);
5089                 }
5090                 spin_unlock_irq(&page->mapping->tree_lock);
5091                 ClearPageError(page);
5092                 unlock_page(page);
5093         }
5094         WARN_ON(atomic_read(&eb->refs) == 0);
5095 }
5096
5097 int set_extent_buffer_dirty(struct extent_buffer *eb)
5098 {
5099         unsigned long i;
5100         unsigned long num_pages;
5101         int was_dirty = 0;
5102
5103         check_buffer_tree_ref(eb);
5104
5105         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5106
5107         num_pages = num_extent_pages(eb->start, eb->len);
5108         WARN_ON(atomic_read(&eb->refs) == 0);
5109         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5110
5111         for (i = 0; i < num_pages; i++)
5112                 set_page_dirty(eb->pages[i]);
5113         return was_dirty;
5114 }
5115
5116 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5117 {
5118         unsigned long i;
5119         struct page *page;
5120         unsigned long num_pages;
5121
5122         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5123         num_pages = num_extent_pages(eb->start, eb->len);
5124         for (i = 0; i < num_pages; i++) {
5125                 page = eb->pages[i];
5126                 if (page)
5127                         ClearPageUptodate(page);
5128         }
5129 }
5130
5131 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5132 {
5133         unsigned long i;
5134         struct page *page;
5135         unsigned long num_pages;
5136
5137         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5138         num_pages = num_extent_pages(eb->start, eb->len);
5139         for (i = 0; i < num_pages; i++) {
5140                 page = eb->pages[i];
5141                 SetPageUptodate(page);
5142         }
5143 }
5144
5145 int extent_buffer_uptodate(struct extent_buffer *eb)
5146 {
5147         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5148 }
5149
5150 int read_extent_buffer_pages(struct extent_io_tree *tree,
5151                              struct extent_buffer *eb, u64 start, int wait,
5152                              get_extent_t *get_extent, int mirror_num)
5153 {
5154         unsigned long i;
5155         unsigned long start_i;
5156         struct page *page;
5157         int err;
5158         int ret = 0;
5159         int locked_pages = 0;
5160         int all_uptodate = 1;
5161         unsigned long num_pages;
5162         unsigned long num_reads = 0;
5163         struct bio *bio = NULL;
5164         unsigned long bio_flags = 0;
5165
5166         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5167                 return 0;
5168
5169         if (start) {
5170                 WARN_ON(start < eb->start);
5171                 start_i = (start >> PAGE_CACHE_SHIFT) -
5172                         (eb->start >> PAGE_CACHE_SHIFT);
5173         } else {
5174                 start_i = 0;
5175         }
5176
5177         num_pages = num_extent_pages(eb->start, eb->len);
5178         for (i = start_i; i < num_pages; i++) {
5179                 page = eb->pages[i];
5180                 if (wait == WAIT_NONE) {
5181                         if (!trylock_page(page))
5182                                 goto unlock_exit;
5183                 } else {
5184                         lock_page(page);
5185                 }
5186                 locked_pages++;
5187                 if (!PageUptodate(page)) {
5188                         num_reads++;
5189                         all_uptodate = 0;
5190                 }
5191         }
5192         if (all_uptodate) {
5193                 if (start_i == 0)
5194                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5195                 goto unlock_exit;
5196         }
5197
5198         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5199         eb->read_mirror = 0;
5200         atomic_set(&eb->io_pages, num_reads);
5201         for (i = start_i; i < num_pages; i++) {
5202                 page = eb->pages[i];
5203                 if (!PageUptodate(page)) {
5204                         ClearPageError(page);
5205                         err = __extent_read_full_page(tree, page,
5206                                                       get_extent, &bio,
5207                                                       mirror_num, &bio_flags,
5208                                                       READ | REQ_META);
5209                         if (err)
5210                                 ret = err;
5211                 } else {
5212                         unlock_page(page);
5213                 }
5214         }
5215
5216         if (bio) {
5217                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
5218                                      bio_flags);
5219                 if (err)
5220                         return err;
5221         }
5222
5223         if (ret || wait != WAIT_COMPLETE)
5224                 return ret;
5225
5226         for (i = start_i; i < num_pages; i++) {
5227                 page = eb->pages[i];
5228                 wait_on_page_locked(page);
5229                 if (!PageUptodate(page))
5230                         ret = -EIO;
5231         }
5232
5233         return ret;
5234
5235 unlock_exit:
5236         i = start_i;
5237         while (locked_pages > 0) {
5238                 page = eb->pages[i];
5239                 i++;
5240                 unlock_page(page);
5241                 locked_pages--;
5242         }
5243         return ret;
5244 }
5245
5246 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5247                         unsigned long start,
5248                         unsigned long len)
5249 {
5250         size_t cur;
5251         size_t offset;
5252         struct page *page;
5253         char *kaddr;
5254         char *dst = (char *)dstv;
5255         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5256         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5257
5258         WARN_ON(start > eb->len);
5259         WARN_ON(start + len > eb->start + eb->len);
5260
5261         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5262
5263         while (len > 0) {
5264                 page = eb->pages[i];
5265
5266                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5267                 kaddr = page_address(page);
5268                 memcpy(dst, kaddr + offset, cur);
5269
5270                 dst += cur;
5271                 len -= cur;
5272                 offset = 0;
5273                 i++;
5274         }
5275 }
5276
5277 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5278                         unsigned long start,
5279                         unsigned long len)
5280 {
5281         size_t cur;
5282         size_t offset;
5283         struct page *page;
5284         char *kaddr;
5285         char __user *dst = (char __user *)dstv;
5286         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5287         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5288         int ret = 0;
5289
5290         WARN_ON(start > eb->len);
5291         WARN_ON(start + len > eb->start + eb->len);
5292
5293         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5294
5295         while (len > 0) {
5296                 page = eb->pages[i];
5297
5298                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5299                 kaddr = page_address(page);
5300                 if (copy_to_user(dst, kaddr + offset, cur)) {
5301                         ret = -EFAULT;
5302                         break;
5303                 }
5304
5305                 dst += cur;
5306                 len -= cur;
5307                 offset = 0;
5308                 i++;
5309         }
5310
5311         return ret;
5312 }
5313
5314 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5315                                unsigned long min_len, char **map,
5316                                unsigned long *map_start,
5317                                unsigned long *map_len)
5318 {
5319         size_t offset = start & (PAGE_CACHE_SIZE - 1);
5320         char *kaddr;
5321         struct page *p;
5322         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5323         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5324         unsigned long end_i = (start_offset + start + min_len - 1) >>
5325                 PAGE_CACHE_SHIFT;
5326
5327         if (i != end_i)
5328                 return -EINVAL;
5329
5330         if (i == 0) {
5331                 offset = start_offset;
5332                 *map_start = 0;
5333         } else {
5334                 offset = 0;
5335                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
5336         }
5337
5338         if (start + min_len > eb->len) {
5339                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
5340                        "wanted %lu %lu\n",
5341                        eb->start, eb->len, start, min_len);
5342                 return -EINVAL;
5343         }
5344
5345         p = eb->pages[i];
5346         kaddr = page_address(p);
5347         *map = kaddr + offset;
5348         *map_len = PAGE_CACHE_SIZE - offset;
5349         return 0;
5350 }
5351
5352 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5353                           unsigned long start,
5354                           unsigned long len)
5355 {
5356         size_t cur;
5357         size_t offset;
5358         struct page *page;
5359         char *kaddr;
5360         char *ptr = (char *)ptrv;
5361         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5362         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5363         int ret = 0;
5364
5365         WARN_ON(start > eb->len);
5366         WARN_ON(start + len > eb->start + eb->len);
5367
5368         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5369
5370         while (len > 0) {
5371                 page = eb->pages[i];
5372
5373                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5374
5375                 kaddr = page_address(page);
5376                 ret = memcmp(ptr, kaddr + offset, cur);
5377                 if (ret)
5378                         break;
5379
5380                 ptr += cur;
5381                 len -= cur;
5382                 offset = 0;
5383                 i++;
5384         }
5385         return ret;
5386 }
5387
5388 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5389                          unsigned long start, unsigned long len)
5390 {
5391         size_t cur;
5392         size_t offset;
5393         struct page *page;
5394         char *kaddr;
5395         char *src = (char *)srcv;
5396         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5397         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5398
5399         WARN_ON(start > eb->len);
5400         WARN_ON(start + len > eb->start + eb->len);
5401
5402         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5403
5404         while (len > 0) {
5405                 page = eb->pages[i];
5406                 WARN_ON(!PageUptodate(page));
5407
5408                 cur = min(len, PAGE_CACHE_SIZE - offset);
5409                 kaddr = page_address(page);
5410                 memcpy(kaddr + offset, src, cur);
5411
5412                 src += cur;
5413                 len -= cur;
5414                 offset = 0;
5415                 i++;
5416         }
5417 }
5418
5419 void memset_extent_buffer(struct extent_buffer *eb, char c,
5420                           unsigned long start, unsigned long len)
5421 {
5422         size_t cur;
5423         size_t offset;
5424         struct page *page;
5425         char *kaddr;
5426         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5427         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5428
5429         WARN_ON(start > eb->len);
5430         WARN_ON(start + len > eb->start + eb->len);
5431
5432         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5433
5434         while (len > 0) {
5435                 page = eb->pages[i];
5436                 WARN_ON(!PageUptodate(page));
5437
5438                 cur = min(len, PAGE_CACHE_SIZE - offset);
5439                 kaddr = page_address(page);
5440                 memset(kaddr + offset, c, cur);
5441
5442                 len -= cur;
5443                 offset = 0;
5444                 i++;
5445         }
5446 }
5447
5448 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5449                         unsigned long dst_offset, unsigned long src_offset,
5450                         unsigned long len)
5451 {
5452         u64 dst_len = dst->len;
5453         size_t cur;
5454         size_t offset;
5455         struct page *page;
5456         char *kaddr;
5457         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5458         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5459
5460         WARN_ON(src->len != dst_len);
5461
5462         offset = (start_offset + dst_offset) &
5463                 (PAGE_CACHE_SIZE - 1);
5464
5465         while (len > 0) {
5466                 page = dst->pages[i];
5467                 WARN_ON(!PageUptodate(page));
5468
5469                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5470
5471                 kaddr = page_address(page);
5472                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5473
5474                 src_offset += cur;
5475                 len -= cur;
5476                 offset = 0;
5477                 i++;
5478         }
5479 }
5480
5481 /*
5482  * The extent buffer bitmap operations are done with byte granularity because
5483  * bitmap items are not guaranteed to be aligned to a word and therefore a
5484  * single word in a bitmap may straddle two pages in the extent buffer.
5485  */
5486 #define BIT_BYTE(nr) ((nr) / BITS_PER_BYTE)
5487 #define BYTE_MASK ((1 << BITS_PER_BYTE) - 1)
5488 #define BITMAP_FIRST_BYTE_MASK(start) \
5489         ((BYTE_MASK << ((start) & (BITS_PER_BYTE - 1))) & BYTE_MASK)
5490 #define BITMAP_LAST_BYTE_MASK(nbits) \
5491         (BYTE_MASK >> (-(nbits) & (BITS_PER_BYTE - 1)))
5492
5493 /*
5494  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5495  * given bit number
5496  * @eb: the extent buffer
5497  * @start: offset of the bitmap item in the extent buffer
5498  * @nr: bit number
5499  * @page_index: return index of the page in the extent buffer that contains the
5500  * given bit number
5501  * @page_offset: return offset into the page given by page_index
5502  *
5503  * This helper hides the ugliness of finding the byte in an extent buffer which
5504  * contains a given bit.
5505  */
5506 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5507                                     unsigned long start, unsigned long nr,
5508                                     unsigned long *page_index,
5509                                     size_t *page_offset)
5510 {
5511         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5512         size_t byte_offset = BIT_BYTE(nr);
5513         size_t offset;
5514
5515         /*
5516          * The byte we want is the offset of the extent buffer + the offset of
5517          * the bitmap item in the extent buffer + the offset of the byte in the
5518          * bitmap item.
5519          */
5520         offset = start_offset + start + byte_offset;
5521
5522         *page_index = offset >> PAGE_CACHE_SHIFT;
5523         *page_offset = offset & (PAGE_CACHE_SIZE - 1);
5524 }
5525
5526 /**
5527  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5528  * @eb: the extent buffer
5529  * @start: offset of the bitmap item in the extent buffer
5530  * @nr: bit number to test
5531  */
5532 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5533                            unsigned long nr)
5534 {
5535         char *kaddr;
5536         struct page *page;
5537         unsigned long i;
5538         size_t offset;
5539
5540         eb_bitmap_offset(eb, start, nr, &i, &offset);
5541         page = eb->pages[i];
5542         WARN_ON(!PageUptodate(page));
5543         kaddr = page_address(page);
5544         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5545 }
5546
5547 /**
5548  * extent_buffer_bitmap_set - set an area of a bitmap
5549  * @eb: the extent buffer
5550  * @start: offset of the bitmap item in the extent buffer
5551  * @pos: bit number of the first bit
5552  * @len: number of bits to set
5553  */
5554 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5555                               unsigned long pos, unsigned long len)
5556 {
5557         char *kaddr;
5558         struct page *page;
5559         unsigned long i;
5560         size_t offset;
5561         const unsigned int size = pos + len;
5562         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5563         unsigned int mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5564
5565         eb_bitmap_offset(eb, start, pos, &i, &offset);
5566         page = eb->pages[i];
5567         WARN_ON(!PageUptodate(page));
5568         kaddr = page_address(page);
5569
5570         while (len >= bits_to_set) {
5571                 kaddr[offset] |= mask_to_set;
5572                 len -= bits_to_set;
5573                 bits_to_set = BITS_PER_BYTE;
5574                 mask_to_set = ~0U;
5575                 if (++offset >= PAGE_CACHE_SIZE && len > 0) {
5576                         offset = 0;
5577                         page = eb->pages[++i];
5578                         WARN_ON(!PageUptodate(page));
5579                         kaddr = page_address(page);
5580                 }
5581         }
5582         if (len) {
5583                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5584                 kaddr[offset] |= mask_to_set;
5585         }
5586 }
5587
5588
5589 /**
5590  * extent_buffer_bitmap_clear - clear an area of a bitmap
5591  * @eb: the extent buffer
5592  * @start: offset of the bitmap item in the extent buffer
5593  * @pos: bit number of the first bit
5594  * @len: number of bits to clear
5595  */
5596 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5597                                 unsigned long pos, unsigned long len)
5598 {
5599         char *kaddr;
5600         struct page *page;
5601         unsigned long i;
5602         size_t offset;
5603         const unsigned int size = pos + len;
5604         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5605         unsigned int mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5606
5607         eb_bitmap_offset(eb, start, pos, &i, &offset);
5608         page = eb->pages[i];
5609         WARN_ON(!PageUptodate(page));
5610         kaddr = page_address(page);
5611
5612         while (len >= bits_to_clear) {
5613                 kaddr[offset] &= ~mask_to_clear;
5614                 len -= bits_to_clear;
5615                 bits_to_clear = BITS_PER_BYTE;
5616                 mask_to_clear = ~0U;
5617                 if (++offset >= PAGE_CACHE_SIZE && len > 0) {
5618                         offset = 0;
5619                         page = eb->pages[++i];
5620                         WARN_ON(!PageUptodate(page));
5621                         kaddr = page_address(page);
5622                 }
5623         }
5624         if (len) {
5625                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5626                 kaddr[offset] &= ~mask_to_clear;
5627         }
5628 }
5629
5630 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5631 {
5632         unsigned long distance = (src > dst) ? src - dst : dst - src;
5633         return distance < len;
5634 }
5635
5636 static void copy_pages(struct page *dst_page, struct page *src_page,
5637                        unsigned long dst_off, unsigned long src_off,
5638                        unsigned long len)
5639 {
5640         char *dst_kaddr = page_address(dst_page);
5641         char *src_kaddr;
5642         int must_memmove = 0;
5643
5644         if (dst_page != src_page) {
5645                 src_kaddr = page_address(src_page);
5646         } else {
5647                 src_kaddr = dst_kaddr;
5648                 if (areas_overlap(src_off, dst_off, len))
5649                         must_memmove = 1;
5650         }
5651
5652         if (must_memmove)
5653                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5654         else
5655                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5656 }
5657
5658 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5659                            unsigned long src_offset, unsigned long len)
5660 {
5661         size_t cur;
5662         size_t dst_off_in_page;
5663         size_t src_off_in_page;
5664         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5665         unsigned long dst_i;
5666         unsigned long src_i;
5667
5668         if (src_offset + len > dst->len) {
5669                 btrfs_err(dst->fs_info,
5670                         "memmove bogus src_offset %lu move "
5671                        "len %lu dst len %lu", src_offset, len, dst->len);
5672                 BUG_ON(1);
5673         }
5674         if (dst_offset + len > dst->len) {
5675                 btrfs_err(dst->fs_info,
5676                         "memmove bogus dst_offset %lu move "
5677                        "len %lu dst len %lu", dst_offset, len, dst->len);
5678                 BUG_ON(1);
5679         }
5680
5681         while (len > 0) {
5682                 dst_off_in_page = (start_offset + dst_offset) &
5683                         (PAGE_CACHE_SIZE - 1);
5684                 src_off_in_page = (start_offset + src_offset) &
5685                         (PAGE_CACHE_SIZE - 1);
5686
5687                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5688                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5689
5690                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5691                                                src_off_in_page));
5692                 cur = min_t(unsigned long, cur,
5693                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5694
5695                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5696                            dst_off_in_page, src_off_in_page, cur);
5697
5698                 src_offset += cur;
5699                 dst_offset += cur;
5700                 len -= cur;
5701         }
5702 }
5703
5704 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5705                            unsigned long src_offset, unsigned long len)
5706 {
5707         size_t cur;
5708         size_t dst_off_in_page;
5709         size_t src_off_in_page;
5710         unsigned long dst_end = dst_offset + len - 1;
5711         unsigned long src_end = src_offset + len - 1;
5712         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5713         unsigned long dst_i;
5714         unsigned long src_i;
5715
5716         if (src_offset + len > dst->len) {
5717                 btrfs_err(dst->fs_info, "memmove bogus src_offset %lu move "
5718                        "len %lu len %lu", src_offset, len, dst->len);
5719                 BUG_ON(1);
5720         }
5721         if (dst_offset + len > dst->len) {
5722                 btrfs_err(dst->fs_info, "memmove bogus dst_offset %lu move "
5723                        "len %lu len %lu", dst_offset, len, dst->len);
5724                 BUG_ON(1);
5725         }
5726         if (dst_offset < src_offset) {
5727                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5728                 return;
5729         }
5730         while (len > 0) {
5731                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5732                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5733
5734                 dst_off_in_page = (start_offset + dst_end) &
5735                         (PAGE_CACHE_SIZE - 1);
5736                 src_off_in_page = (start_offset + src_end) &
5737                         (PAGE_CACHE_SIZE - 1);
5738
5739                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5740                 cur = min(cur, dst_off_in_page + 1);
5741                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5742                            dst_off_in_page - cur + 1,
5743                            src_off_in_page - cur + 1, cur);
5744
5745                 dst_end -= cur;
5746                 src_end -= cur;
5747                 len -= cur;
5748         }
5749 }
5750
5751 int try_release_extent_buffer(struct page *page)
5752 {
5753         struct extent_buffer *eb;
5754
5755         /*
5756          * We need to make sure noboody is attaching this page to an eb right
5757          * now.
5758          */
5759         spin_lock(&page->mapping->private_lock);
5760         if (!PagePrivate(page)) {
5761                 spin_unlock(&page->mapping->private_lock);
5762                 return 1;
5763         }
5764
5765         eb = (struct extent_buffer *)page->private;
5766         BUG_ON(!eb);
5767
5768         /*
5769          * This is a little awful but should be ok, we need to make sure that
5770          * the eb doesn't disappear out from under us while we're looking at
5771          * this page.
5772          */
5773         spin_lock(&eb->refs_lock);
5774         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5775                 spin_unlock(&eb->refs_lock);
5776                 spin_unlock(&page->mapping->private_lock);
5777                 return 0;
5778         }
5779         spin_unlock(&page->mapping->private_lock);
5780
5781         /*
5782          * If tree ref isn't set then we know the ref on this eb is a real ref,
5783          * so just return, this page will likely be freed soon anyway.
5784          */
5785         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5786                 spin_unlock(&eb->refs_lock);
5787                 return 0;
5788         }
5789
5790         return release_extent_buffer(eb);
5791 }