fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/spinlock.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/swap.h>
  10 #include <linux/writeback.h>
  11 #include <linux/pagevec.h>
  12 #include <linux/prefetch.h>
  13 #include <linux/cleancache.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "ctree.h"
  17 #include "btrfs_inode.h"
  18 #include "volumes.h"
  19 #include "check-integrity.h"
  20 #include "locking.h"
  21 #include "rcu-string.h"
  22 #include "backref.h"
  23 #include "transaction.h"
  24
  25 static struct kmem_cache *extent_state_cache;
  26 static struct kmem_cache *extent_buffer_cache;
  27 static struct bio_set *btrfs_bioset;
  28
  29 static inline bool extent_state_in_tree(const struct extent_state *state)
  30 {
  31         return !RB_EMPTY_NODE(&state->rb_node);
  32 }
  33
  34 #ifdef CONFIG_BTRFS_DEBUG
  35 static LIST_HEAD(buffers);
  36 static LIST_HEAD(states);
  37
  38 static DEFINE_SPINLOCK(leak_lock);
  39
  40 static inline
  41 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  42 {
  43         unsigned long flags;
  44
  45         spin_lock_irqsave(&leak_lock, flags);
  46         list_add(new, head);
  47         spin_unlock_irqrestore(&leak_lock, flags);
  48 }
  49
  50 static inline
  51 void btrfs_leak_debug_del(struct list_head *entry)
  52 {
  53         unsigned long flags;
  54
  55         spin_lock_irqsave(&leak_lock, flags);
  56         list_del(entry);
  57         spin_unlock_irqrestore(&leak_lock, flags);
  58 }
  59
  60 static inline
  61 void btrfs_leak_debug_check(void)
  62 {
  63         struct extent_state *state;
  64         struct extent_buffer *eb;
  65
  66         while (!list_empty(&states)) {
  67                 state = list_entry(states.next, struct extent_state, leak_list);
  68                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  69                        state->start, state->end, state->state,
  70                        extent_state_in_tree(state),
  71                        atomic_read(&state->refs));
  72                 list_del(&state->leak_list);
  73                 kmem_cache_free(extent_state_cache, state);
  74         }
  75
  76         while (!list_empty(&buffers)) {
  77                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  78                 printk(KERN_ERR "BTRFS: buffer leak start %llu len %lu refs %d\n",
  79                        eb->start, eb->len, atomic_read(&eb->refs));
  80                 list_del(&eb->leak_list);
  81                 kmem_cache_free(extent_buffer_cache, eb);
  82         }
  83 }
  84
  85 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  86         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  87 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  88                 struct extent_io_tree *tree, u64 start, u64 end)
  89 {
  90         struct inode *inode;
  91         u64 isize;
  92
  93         if (!tree->mapping)
  94                 return;
  95
  96         inode = tree->mapping->host;
  97         isize = i_size_read(inode);
  98         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
  99                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
 100                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
 101                                 caller, btrfs_ino(inode), isize, start, end);
 102         }
 103 }
 104 #else
 105 #define btrfs_leak_debug_add(new, head) do {} while (0)
 106 #define btrfs_leak_debug_del(entry)     do {} while (0)
 107 #define btrfs_leak_debug_check()        do {} while (0)
 108 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 109 #endif
 110
 111 #define BUFFER_LRU_MAX 64
 112
 113 struct tree_entry {
 114         u64 start;
 115         u64 end;
 116         struct rb_node rb_node;
 117 };
 118
 119 struct extent_page_data {
 120         struct bio *bio;
 121         struct extent_io_tree *tree;
 122         get_extent_t *get_extent;
 123         unsigned long bio_flags;
 124
 125         /* tells writepage not to lock the state bits for this range
 126          * it still does the unlocking
 127          */
 128         unsigned int extent_locked:1;
 129
 130         /* tells the submit_bio code to use a WRITE_SYNC */
 131         unsigned int sync_io:1;
 132 };
 133
 134 static void add_extent_changeset(struct extent_state *state, unsigned bits,
 135                                  struct extent_changeset *changeset,
 136                                  int set)
 137 {
 138         int ret;
 139
 140         if (!changeset)
 141                 return;
 142         if (set && (state->state & bits) == bits)
 143                 return;
 144         if (!set && (state->state & bits) == 0)
 145                 return;
 146         changeset->bytes_changed += state->end - state->start + 1;
 147         ret = ulist_add(changeset->range_changed, state->start, state->end,
 148                         GFP_ATOMIC);
 149         /* ENOMEM */
 150         BUG_ON(ret < 0);
 151 }
 152
 153 static noinline void flush_write_bio(void *data);
 154 static inline struct btrfs_fs_info *
 155 tree_fs_info(struct extent_io_tree *tree)
 156 {
 157         if (!tree->mapping)
 158                 return NULL;
 159         return btrfs_sb(tree->mapping->host->i_sb);
 160 }
 161
 162 int __init extent_io_init(void)
 163 {
 164         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 165                         sizeof(struct extent_state), 0,
 166                         SLAB_MEM_SPREAD, NULL);
 167         if (!extent_state_cache)
 168                 return -ENOMEM;
 169
 170         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 171                         sizeof(struct extent_buffer), 0,
 172                         SLAB_MEM_SPREAD, NULL);
 173         if (!extent_buffer_cache)
 174                 goto free_state_cache;
 175
 176         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 177                                      offsetof(struct btrfs_io_bio, bio));
 178         if (!btrfs_bioset)
 179                 goto free_buffer_cache;
 180
 181         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 182                 goto free_bioset;
 183
 184         return 0;
 185
 186 free_bioset:
 187         bioset_free(btrfs_bioset);
 188         btrfs_bioset = NULL;
 189
 190 free_buffer_cache:
 191         kmem_cache_destroy(extent_buffer_cache);
 192         extent_buffer_cache = NULL;
 193
 194 free_state_cache:
 195         kmem_cache_destroy(extent_state_cache);
 196         extent_state_cache = NULL;
 197         return -ENOMEM;
 198 }
 199
 200 void extent_io_exit(void)
 201 {
 202         btrfs_leak_debug_check();
 203
 204         /*
 205          * Make sure all delayed rcu free are flushed before we
 206          * destroy caches.
 207          */
 208         rcu_barrier();
 209         kmem_cache_destroy(extent_state_cache);
 210         kmem_cache_destroy(extent_buffer_cache);
 211         if (btrfs_bioset)
 212                 bioset_free(btrfs_bioset);
 213 }
 214
 215 void extent_io_tree_init(struct extent_io_tree *tree,
 216                          struct address_space *mapping)
 217 {
 218         tree->state = RB_ROOT;
 219         tree->ops = NULL;
 220         tree->dirty_bytes = 0;
 221         spin_lock_init(&tree->lock);
 222         tree->mapping = mapping;
 223 }
 224
 225 static struct extent_state *alloc_extent_state(gfp_t mask)
 226 {
 227         struct extent_state *state;
 228
 229         state = kmem_cache_alloc(extent_state_cache, mask);
 230         if (!state)
 231                 return state;
 232         state->state = 0;
 233         state->failrec = NULL;
 234         RB_CLEAR_NODE(&state->rb_node);
 235         btrfs_leak_debug_add(&state->leak_list, &states);
 236         atomic_set(&state->refs, 1);
 237         init_waitqueue_head(&state->wq);
 238         trace_alloc_extent_state(state, mask, _RET_IP_);
 239         return state;
 240 }
 241
 242 void free_extent_state(struct extent_state *state)
 243 {
 244         if (!state)
 245                 return;
 246         if (atomic_dec_and_test(&state->refs)) {
 247                 WARN_ON(extent_state_in_tree(state));
 248                 btrfs_leak_debug_del(&state->leak_list);
 249                 trace_free_extent_state(state, _RET_IP_);
 250                 kmem_cache_free(extent_state_cache, state);
 251         }
 252 }
 253
 254 static struct rb_node *tree_insert(struct rb_root *root,
 255                                    struct rb_node *search_start,
 256                                    u64 offset,
 257                                    struct rb_node *node,
 258                                    struct rb_node ***p_in,
 259                                    struct rb_node **parent_in)
 260 {
 261         struct rb_node **p;
 262         struct rb_node *parent = NULL;
 263         struct tree_entry *entry;
 264
 265         if (p_in && parent_in) {
 266                 p = *p_in;
 267                 parent = *parent_in;
 268                 goto do_insert;
 269         }
 270
 271         p = search_start ? &search_start : &root->rb_node;
 272         while (*p) {
 273                 parent = *p;
 274                 entry = rb_entry(parent, struct tree_entry, rb_node);
 275
 276                 if (offset < entry->start)
 277                         p = &(*p)->rb_left;
 278                 else if (offset > entry->end)
 279                         p = &(*p)->rb_right;
 280                 else
 281                         return parent;
 282         }
 283
 284 do_insert:
 285         rb_link_node(node, parent, p);
 286         rb_insert_color(node, root);
 287         return NULL;
 288 }
 289
 290 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 291                                       struct rb_node **prev_ret,
 292                                       struct rb_node **next_ret,
 293                                       struct rb_node ***p_ret,
 294                                       struct rb_node **parent_ret)
 295 {
 296         struct rb_root *root = &tree->state;
 297         struct rb_node **n = &root->rb_node;
 298         struct rb_node *prev = NULL;
 299         struct rb_node *orig_prev = NULL;
 300         struct tree_entry *entry;
 301         struct tree_entry *prev_entry = NULL;
 302
 303         while (*n) {
 304                 prev = *n;
 305                 entry = rb_entry(prev, struct tree_entry, rb_node);
 306                 prev_entry = entry;
 307
 308                 if (offset < entry->start)
 309                         n = &(*n)->rb_left;
 310                 else if (offset > entry->end)
 311                         n = &(*n)->rb_right;
 312                 else
 313                         return *n;
 314         }
 315
 316         if (p_ret)
 317                 *p_ret = n;
 318         if (parent_ret)
 319                 *parent_ret = prev;
 320
 321         if (prev_ret) {
 322                 orig_prev = prev;
 323                 while (prev && offset > prev_entry->end) {
 324                         prev = rb_next(prev);
 325                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 326                 }
 327                 *prev_ret = prev;
 328                 prev = orig_prev;
 329         }
 330
 331         if (next_ret) {
 332                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 333                 while (prev && offset < prev_entry->start) {
 334                         prev = rb_prev(prev);
 335                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 336                 }
 337                 *next_ret = prev;
 338         }
 339         return NULL;
 340 }
 341
 342 static inline struct rb_node *
 343 tree_search_for_insert(struct extent_io_tree *tree,
 344                        u64 offset,
 345                        struct rb_node ***p_ret,
 346                        struct rb_node **parent_ret)
 347 {
 348         struct rb_node *prev = NULL;
 349         struct rb_node *ret;
 350
 351         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 352         if (!ret)
 353                 return prev;
 354         return ret;
 355 }
 356
 357 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 358                                           u64 offset)
 359 {
 360         return tree_search_for_insert(tree, offset, NULL, NULL);
 361 }
 362
 363 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 364                      struct extent_state *other)
 365 {
 366         if (tree->ops && tree->ops->merge_extent_hook)
 367                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 368                                              other);
 369 }
 370
 371 /*
 372  * utility function to look for merge candidates inside a given range.
 373  * Any extents with matching state are merged together into a single
 374  * extent in the tree.  Extents with EXTENT_IO in their state field
 375  * are not merged because the end_io handlers need to be able to do
 376  * operations on them without sleeping (or doing allocations/splits).
 377  *
 378  * This should be called with the tree lock held.
 379  */
 380 static void merge_state(struct extent_io_tree *tree,
 381                         struct extent_state *state)
 382 {
 383         struct extent_state *other;
 384         struct rb_node *other_node;
 385
 386         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 387                 return;
 388
 389         other_node = rb_prev(&state->rb_node);
 390         if (other_node) {
 391                 other = rb_entry(other_node, struct extent_state, rb_node);
 392                 if (other->end == state->start - 1 &&
 393                     other->state == state->state) {
 394                         merge_cb(tree, state, other);
 395                         state->start = other->start;
 396                         rb_erase(&other->rb_node, &tree->state);
 397                         RB_CLEAR_NODE(&other->rb_node);
 398                         free_extent_state(other);
 399                 }
 400         }
 401         other_node = rb_next(&state->rb_node);
 402         if (other_node) {
 403                 other = rb_entry(other_node, struct extent_state, rb_node);
 404                 if (other->start == state->end + 1 &&
 405                     other->state == state->state) {
 406                         merge_cb(tree, state, other);
 407                         state->end = other->end;
 408                         rb_erase(&other->rb_node, &tree->state);
 409                         RB_CLEAR_NODE(&other->rb_node);
 410                         free_extent_state(other);
 411                 }
 412         }
 413 }
 414
 415 static void set_state_cb(struct extent_io_tree *tree,
 416                          struct extent_state *state, unsigned *bits)
 417 {
 418         if (tree->ops && tree->ops->set_bit_hook)
 419                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 420 }
 421
 422 static void clear_state_cb(struct extent_io_tree *tree,
 423                            struct extent_state *state, unsigned *bits)
 424 {
 425         if (tree->ops && tree->ops->clear_bit_hook)
 426                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 427 }
 428
 429 static void set_state_bits(struct extent_io_tree *tree,
 430                            struct extent_state *state, unsigned *bits,
 431                            struct extent_changeset *changeset);
 432
 433 /*
 434  * insert an extent_state struct into the tree.  'bits' are set on the
 435  * struct before it is inserted.
 436  *
 437  * This may return -EEXIST if the extent is already there, in which case the
 438  * state struct is freed.
 439  *
 440  * The tree lock is not taken internally.  This is a utility function and
 441  * probably isn't what you want to call (see set/clear_extent_bit).
 442  */
 443 static int insert_state(struct extent_io_tree *tree,
 444                         struct extent_state *state, u64 start, u64 end,
 445                         struct rb_node ***p,
 446                         struct rb_node **parent,
 447                         unsigned *bits, struct extent_changeset *changeset)
 448 {
 449         struct rb_node *node;
 450
 451         if (end < start)
 452                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 453                        end, start);
 454         state->start = start;
 455         state->end = end;
 456
 457         set_state_bits(tree, state, bits, changeset);
 458
 459         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 460         if (node) {
 461                 struct extent_state *found;
 462                 found = rb_entry(node, struct extent_state, rb_node);
 463                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of %llu %llu\n",
 464                        found->start, found->end, start, end);
 465                 return -EEXIST;
 466         }
 467         merge_state(tree, state);
 468         return 0;
 469 }
 470
 471 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 472                      u64 split)
 473 {
 474         if (tree->ops && tree->ops->split_extent_hook)
 475                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 476 }
 477
 478 /*
 479  * split a given extent state struct in two, inserting the preallocated
 480  * struct 'prealloc' as the newly created second half.  'split' indicates an
 481  * offset inside 'orig' where it should be split.
 482  *
 483  * Before calling,
 484  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 485  * are two extent state structs in the tree:
 486  * prealloc: [orig->start, split - 1]
 487  * orig: [ split, orig->end ]
 488  *
 489  * The tree locks are not taken by this function. They need to be held
 490  * by the caller.
 491  */
 492 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 493                        struct extent_state *prealloc, u64 split)
 494 {
 495         struct rb_node *node;
 496
 497         split_cb(tree, orig, split);
 498
 499         prealloc->start = orig->start;
 500         prealloc->end = split - 1;
 501         prealloc->state = orig->state;
 502         orig->start = split;
 503
 504         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 505                            &prealloc->rb_node, NULL, NULL);
 506         if (node) {
 507                 free_extent_state(prealloc);
 508                 return -EEXIST;
 509         }
 510         return 0;
 511 }
 512
 513 static struct extent_state *next_state(struct extent_state *state)
 514 {
 515         struct rb_node *next = rb_next(&state->rb_node);
 516         if (next)
 517                 return rb_entry(next, struct extent_state, rb_node);
 518         else
 519                 return NULL;
 520 }
 521
 522 /*
 523  * utility function to clear some bits in an extent state struct.
 524  * it will optionally wake up any one waiting on this state (wake == 1).
 525  *
 526  * If no bits are set on the state struct after clearing things, the
 527  * struct is freed and removed from the tree
 528  */
 529 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 530                                             struct extent_state *state,
 531                                             unsigned *bits, int wake,
 532                                             struct extent_changeset *changeset)
 533 {
 534         struct extent_state *next;
 535         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 536
 537         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 538                 u64 range = state->end - state->start + 1;
 539                 WARN_ON(range > tree->dirty_bytes);
 540                 tree->dirty_bytes -= range;
 541         }
 542         clear_state_cb(tree, state, bits);
 543         add_extent_changeset(state, bits_to_clear, changeset, 0);
 544         state->state &= ~bits_to_clear;
 545         if (wake)
 546                 wake_up(&state->wq);
 547         if (state->state == 0) {
 548                 next = next_state(state);
 549                 if (extent_state_in_tree(state)) {
 550                         rb_erase(&state->rb_node, &tree->state);
 551                         RB_CLEAR_NODE(&state->rb_node);
 552                         free_extent_state(state);
 553                 } else {
 554                         WARN_ON(1);
 555                 }
 556         } else {
 557                 merge_state(tree, state);
 558                 next = next_state(state);
 559         }
 560         return next;
 561 }
 562
 563 static struct extent_state *
 564 alloc_extent_state_atomic(struct extent_state *prealloc)
 565 {
 566         if (!prealloc)
 567                 prealloc = alloc_extent_state(GFP_ATOMIC);
 568
 569         return prealloc;
 570 }
 571
 572 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 573 {
 574         btrfs_panic(tree_fs_info(tree), err,
 575                     "Locking error: Extent tree was modified by another thread while locked.");
 576 }
 577
 578 /*
 579  * clear some bits on a range in the tree.  This may require splitting
 580  * or inserting elements in the tree, so the gfp mask is used to
 581  * indicate which allocations or sleeping are allowed.
 582  *
 583  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 584  * the given range from the tree regardless of state (ie for truncate).
 585  *
 586  * the range [start, end] is inclusive.
 587  *
 588  * This takes the tree lock, and returns 0 on success and < 0 on error.
 589  */
 590 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 591                               unsigned bits, int wake, int delete,
 592                               struct extent_state **cached_state,
 593                               gfp_t mask, struct extent_changeset *changeset)
 594 {
 595         struct extent_state *state;
 596         struct extent_state *cached;
 597         struct extent_state *prealloc = NULL;
 598         struct rb_node *node;
 599         u64 last_end;
 600         int err;
 601         int clear = 0;
 602
 603         btrfs_debug_check_extent_io_range(tree, start, end);
 604
 605         if (bits & EXTENT_DELALLOC)
 606                 bits |= EXTENT_NORESERVE;
 607
 608         if (delete)
 609                 bits |= ~EXTENT_CTLBITS;
 610         bits |= EXTENT_FIRST_DELALLOC;
 611
 612         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 613                 clear = 1;
 614 again:
 615         if (!prealloc && gfpflags_allow_blocking(mask)) {
 616                 /*
 617                  * Don't care for allocation failure here because we might end
 618                  * up not needing the pre-allocated extent state at all, which
 619                  * is the case if we only have in the tree extent states that
 620                  * cover our input range and don't cover too any other range.
 621                  * If we end up needing a new extent state we allocate it later.
 622                  */
 623                 prealloc = alloc_extent_state(mask);
 624         }
 625
 626         spin_lock(&tree->lock);
 627         if (cached_state) {
 628                 cached = *cached_state;
 629
 630                 if (clear) {
 631                         *cached_state = NULL;
 632                         cached_state = NULL;
 633                 }
 634
 635                 if (cached && extent_state_in_tree(cached) &&
 636                     cached->start <= start && cached->end > start) {
 637                         if (clear)
 638                                 atomic_dec(&cached->refs);
 639                         state = cached;
 640                         goto hit_next;
 641                 }
 642                 if (clear)
 643                         free_extent_state(cached);
 644         }
 645         /*
 646          * this search will find the extents that end after
 647          * our range starts
 648          */
 649         node = tree_search(tree, start);
 650         if (!node)
 651                 goto out;
 652         state = rb_entry(node, struct extent_state, rb_node);
 653 hit_next:
 654         if (state->start > end)
 655                 goto out;
 656         WARN_ON(state->end < start);
 657         last_end = state->end;
 658
 659         /* the state doesn't have the wanted bits, go ahead */
 660         if (!(state->state & bits)) {
 661                 state = next_state(state);
 662                 goto next;
 663         }
 664
 665         /*
 666          *     | ---- desired range ---- |
 667          *  | state | or
 668          *  | ------------- state -------------- |
 669          *
 670          * We need to split the extent we found, and may flip
 671          * bits on second half.
 672          *
 673          * If the extent we found extends past our range, we
 674          * just split and search again.  It'll get split again
 675          * the next time though.
 676          *
 677          * If the extent we found is inside our range, we clear
 678          * the desired bit on it.
 679          */
 680
 681         if (state->start < start) {
 682                 prealloc = alloc_extent_state_atomic(prealloc);
 683                 BUG_ON(!prealloc);
 684                 err = split_state(tree, state, prealloc, start);
 685                 if (err)
 686                         extent_io_tree_panic(tree, err);
 687
 688                 prealloc = NULL;
 689                 if (err)
 690                         goto out;
 691                 if (state->end <= end) {
 692                         state = clear_state_bit(tree, state, &bits, wake,
 693                                                 changeset);
 694                         goto next;
 695                 }
 696                 goto search_again;
 697         }
 698         /*
 699          * | ---- desired range ---- |
 700          *                        | state |
 701          * We need to split the extent, and clear the bit
 702          * on the first half
 703          */
 704         if (state->start <= end && state->end > end) {
 705                 prealloc = alloc_extent_state_atomic(prealloc);
 706                 BUG_ON(!prealloc);
 707                 err = split_state(tree, state, prealloc, end + 1);
 708                 if (err)
 709                         extent_io_tree_panic(tree, err);
 710
 711                 if (wake)
 712                         wake_up(&state->wq);
 713
 714                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 715
 716                 prealloc = NULL;
 717                 goto out;
 718         }
 719
 720         state = clear_state_bit(tree, state, &bits, wake, changeset);
 721 next:
 722         if (last_end == (u64)-1)
 723                 goto out;
 724         start = last_end + 1;
 725         if (start <= end && state && !need_resched())
 726                 goto hit_next;
 727
 728 search_again:
 729         if (start > end)
 730                 goto out;
 731         spin_unlock(&tree->lock);
 732         if (gfpflags_allow_blocking(mask))
 733                 cond_resched();
 734         goto again;
 735
 736 out:
 737         spin_unlock(&tree->lock);
 738         if (prealloc)
 739                 free_extent_state(prealloc);
 740
 741         return 0;
 742
 743 }
 744
 745 static void wait_on_state(struct extent_io_tree *tree,
 746                           struct extent_state *state)
 747                 __releases(tree->lock)
 748                 __acquires(tree->lock)
 749 {
 750         DEFINE_WAIT(wait);
 751         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 752         spin_unlock(&tree->lock);
 753         schedule();
 754         spin_lock(&tree->lock);
 755         finish_wait(&state->wq, &wait);
 756 }
 757
 758 /*
 759  * waits for one or more bits to clear on a range in the state tree.
 760  * The range [start, end] is inclusive.
 761  * The tree lock is taken by this function
 762  */
 763 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 764                             unsigned long bits)
 765 {
 766         struct extent_state *state;
 767         struct rb_node *node;
 768
 769         btrfs_debug_check_extent_io_range(tree, start, end);
 770
 771         spin_lock(&tree->lock);
 772 again:
 773         while (1) {
 774                 /*
 775                  * this search will find all the extents that end after
 776                  * our range starts
 777                  */
 778                 node = tree_search(tree, start);
 779 process_node:
 780                 if (!node)
 781                         break;
 782
 783                 state = rb_entry(node, struct extent_state, rb_node);
 784
 785                 if (state->start > end)
 786                         goto out;
 787
 788                 if (state->state & bits) {
 789                         start = state->start;
 790                         atomic_inc(&state->refs);
 791                         wait_on_state(tree, state);
 792                         free_extent_state(state);
 793                         goto again;
 794                 }
 795                 start = state->end + 1;
 796
 797                 if (start > end)
 798                         break;
 799
 800                 if (!cond_resched_lock(&tree->lock)) {
 801                         node = rb_next(node);
 802                         goto process_node;
 803                 }
 804         }
 805 out:
 806         spin_unlock(&tree->lock);
 807 }
 808
 809 static void set_state_bits(struct extent_io_tree *tree,
 810                            struct extent_state *state,
 811                            unsigned *bits, struct extent_changeset *changeset)
 812 {
 813         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 814
 815         set_state_cb(tree, state, bits);
 816         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 817                 u64 range = state->end - state->start + 1;
 818                 tree->dirty_bytes += range;
 819         }
 820         add_extent_changeset(state, bits_to_set, changeset, 1);
 821         state->state |= bits_to_set;
 822 }
 823
 824 static void cache_state_if_flags(struct extent_state *state,
 825                                  struct extent_state **cached_ptr,
 826                                  unsigned flags)
 827 {
 828         if (cached_ptr && !(*cached_ptr)) {
 829                 if (!flags || (state->state & flags)) {
 830                         *cached_ptr = state;
 831                         atomic_inc(&state->refs);
 832                 }
 833         }
 834 }
 835
 836 static void cache_state(struct extent_state *state,
 837                         struct extent_state **cached_ptr)
 838 {
 839         return cache_state_if_flags(state, cached_ptr,
 840                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 841 }
 842
 843 /*
 844  * set some bits on a range in the tree.  This may require allocations or
 845  * sleeping, so the gfp mask is used to indicate what is allowed.
 846  *
 847  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 848  * part of the range already has the desired bits set.  The start of the
 849  * existing range is returned in failed_start in this case.
 850  *
 851  * [start, end] is inclusive This takes the tree lock.
 852  */
 853
 854 static int __must_check
 855 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 856                  unsigned bits, unsigned exclusive_bits,
 857                  u64 *failed_start, struct extent_state **cached_state,
 858                  gfp_t mask, struct extent_changeset *changeset)
 859 {
 860         struct extent_state *state;
 861         struct extent_state *prealloc = NULL;
 862         struct rb_node *node;
 863         struct rb_node **p;
 864         struct rb_node *parent;
 865         int err = 0;
 866         u64 last_start;
 867         u64 last_end;
 868
 869         btrfs_debug_check_extent_io_range(tree, start, end);
 870
 871         bits |= EXTENT_FIRST_DELALLOC;
 872 again:
 873         if (!prealloc && gfpflags_allow_blocking(mask)) {
 874                 /*
 875                  * Don't care for allocation failure here because we might end
 876                  * up not needing the pre-allocated extent state at all, which
 877                  * is the case if we only have in the tree extent states that
 878                  * cover our input range and don't cover too any other range.
 879                  * If we end up needing a new extent state we allocate it later.
 880                  */
 881                 prealloc = alloc_extent_state(mask);
 882         }
 883
 884         spin_lock(&tree->lock);
 885         if (cached_state && *cached_state) {
 886                 state = *cached_state;
 887                 if (state->start <= start && state->end > start &&
 888                     extent_state_in_tree(state)) {
 889                         node = &state->rb_node;
 890                         goto hit_next;
 891                 }
 892         }
 893         /*
 894          * this search will find all the extents that end after
 895          * our range starts.
 896          */
 897         node = tree_search_for_insert(tree, start, &p, &parent);
 898         if (!node) {
 899                 prealloc = alloc_extent_state_atomic(prealloc);
 900                 BUG_ON(!prealloc);
 901                 err = insert_state(tree, prealloc, start, end,
 902                                    &p, &parent, &bits, changeset);
 903                 if (err)
 904                         extent_io_tree_panic(tree, err);
 905
 906                 cache_state(prealloc, cached_state);
 907                 prealloc = NULL;
 908                 goto out;
 909         }
 910         state = rb_entry(node, struct extent_state, rb_node);
 911 hit_next:
 912         last_start = state->start;
 913         last_end = state->end;
 914
 915         /*
 916          * | ---- desired range ---- |
 917          * | state |
 918          *
 919          * Just lock what we found and keep going
 920          */
 921         if (state->start == start && state->end <= end) {
 922                 if (state->state & exclusive_bits) {
 923                         *failed_start = state->start;
 924                         err = -EEXIST;
 925                         goto out;
 926                 }
 927
 928                 set_state_bits(tree, state, &bits, changeset);
 929                 cache_state(state, cached_state);
 930                 merge_state(tree, state);
 931                 if (last_end == (u64)-1)
 932                         goto out;
 933                 start = last_end + 1;
 934                 state = next_state(state);
 935                 if (start < end && state && state->start == start &&
 936                     !need_resched())
 937                         goto hit_next;
 938                 goto search_again;
 939         }
 940
 941         /*
 942          *     | ---- desired range ---- |
 943          * | state |
 944          *   or
 945          * | ------------- state -------------- |
 946          *
 947          * We need to split the extent we found, and may flip bits on
 948          * second half.
 949          *
 950          * If the extent we found extends past our
 951          * range, we just split and search again.  It'll get split
 952          * again the next time though.
 953          *
 954          * If the extent we found is inside our range, we set the
 955          * desired bit on it.
 956          */
 957         if (state->start < start) {
 958                 if (state->state & exclusive_bits) {
 959                         *failed_start = start;
 960                         err = -EEXIST;
 961                         goto out;
 962                 }
 963
 964                 prealloc = alloc_extent_state_atomic(prealloc);
 965                 BUG_ON(!prealloc);
 966                 err = split_state(tree, state, prealloc, start);
 967                 if (err)
 968                         extent_io_tree_panic(tree, err);
 969
 970                 prealloc = NULL;
 971                 if (err)
 972                         goto out;
 973                 if (state->end <= end) {
 974                         set_state_bits(tree, state, &bits, changeset);
 975                         cache_state(state, cached_state);
 976                         merge_state(tree, state);
 977                         if (last_end == (u64)-1)
 978                                 goto out;
 979                         start = last_end + 1;
 980                         state = next_state(state);
 981                         if (start < end && state && state->start == start &&
 982                             !need_resched())
 983                                 goto hit_next;
 984                 }
 985                 goto search_again;
 986         }
 987         /*
 988          * | ---- desired range ---- |
 989          *     | state | or               | state |
 990          *
 991          * There's a hole, we need to insert something in it and
 992          * ignore the extent we found.
 993          */
 994         if (state->start > start) {
 995                 u64 this_end;
 996                 if (end < last_start)
 997                         this_end = end;
 998                 else
 999                         this_end = last_start - 1;
1000
1001                 prealloc = alloc_extent_state_atomic(prealloc);
1002                 BUG_ON(!prealloc);
1003
1004                 /*
1005                  * Avoid to free 'prealloc' if it can be merged with
1006                  * the later extent.
1007                  */
1008                 err = insert_state(tree, prealloc, start, this_end,
1009                                    NULL, NULL, &bits, changeset);
1010                 if (err)
1011                         extent_io_tree_panic(tree, err);
1012
1013                 cache_state(prealloc, cached_state);
1014                 prealloc = NULL;
1015                 start = this_end + 1;
1016                 goto search_again;
1017         }
1018         /*
1019          * | ---- desired range ---- |
1020          *                        | state |
1021          * We need to split the extent, and set the bit
1022          * on the first half
1023          */
1024         if (state->start <= end && state->end > end) {
1025                 if (state->state & exclusive_bits) {
1026                         *failed_start = start;
1027                         err = -EEXIST;
1028                         goto out;
1029                 }
1030
1031                 prealloc = alloc_extent_state_atomic(prealloc);
1032                 BUG_ON(!prealloc);
1033                 err = split_state(tree, state, prealloc, end + 1);
1034                 if (err)
1035                         extent_io_tree_panic(tree, err);
1036
1037                 set_state_bits(tree, prealloc, &bits, changeset);
1038                 cache_state(prealloc, cached_state);
1039                 merge_state(tree, prealloc);
1040                 prealloc = NULL;
1041                 goto out;
1042         }
1043
1044 search_again:
1045         if (start > end)
1046                 goto out;
1047         spin_unlock(&tree->lock);
1048         if (gfpflags_allow_blocking(mask))
1049                 cond_resched();
1050         goto again;
1051
1052 out:
1053         spin_unlock(&tree->lock);
1054         if (prealloc)
1055                 free_extent_state(prealloc);
1056
1057         return err;
1058
1059 }
1060
1061 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1062                    unsigned bits, u64 * failed_start,
1063                    struct extent_state **cached_state, gfp_t mask)
1064 {
1065         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1066                                 cached_state, mask, NULL);
1067 }
1068
1069
1070 /**
1071  * convert_extent_bit - convert all bits in a given range from one bit to
1072  *                      another
1073  * @tree:       the io tree to search
1074  * @start:      the start offset in bytes
1075  * @end:        the end offset in bytes (inclusive)
1076  * @bits:       the bits to set in this range
1077  * @clear_bits: the bits to clear in this range
1078  * @cached_state:       state that we're going to cache
1079  *
1080  * This will go through and set bits for the given range.  If any states exist
1081  * already in this range they are set with the given bit and cleared of the
1082  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1083  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1084  * boundary bits like LOCK.
1085  *
1086  * All allocations are done with GFP_NOFS.
1087  */
1088 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1089                        unsigned bits, unsigned clear_bits,
1090                        struct extent_state **cached_state)
1091 {
1092         struct extent_state *state;
1093         struct extent_state *prealloc = NULL;
1094         struct rb_node *node;
1095         struct rb_node **p;
1096         struct rb_node *parent;
1097         int err = 0;
1098         u64 last_start;
1099         u64 last_end;
1100         bool first_iteration = true;
1101
1102         btrfs_debug_check_extent_io_range(tree, start, end);
1103
1104 again:
1105         if (!prealloc) {
1106                 /*
1107                  * Best effort, don't worry if extent state allocation fails
1108                  * here for the first iteration. We might have a cached state
1109                  * that matches exactly the target range, in which case no
1110                  * extent state allocations are needed. We'll only know this
1111                  * after locking the tree.
1112                  */
1113                 prealloc = alloc_extent_state(GFP_NOFS);
1114                 if (!prealloc && !first_iteration)
1115                         return -ENOMEM;
1116         }
1117
1118         spin_lock(&tree->lock);
1119         if (cached_state && *cached_state) {
1120                 state = *cached_state;
1121                 if (state->start <= start && state->end > start &&
1122                     extent_state_in_tree(state)) {
1123                         node = &state->rb_node;
1124                         goto hit_next;
1125                 }
1126         }
1127
1128         /*
1129          * this search will find all the extents that end after
1130          * our range starts.
1131          */
1132         node = tree_search_for_insert(tree, start, &p, &parent);
1133         if (!node) {
1134                 prealloc = alloc_extent_state_atomic(prealloc);
1135                 if (!prealloc) {
1136                         err = -ENOMEM;
1137                         goto out;
1138                 }
1139                 err = insert_state(tree, prealloc, start, end,
1140                                    &p, &parent, &bits, NULL);
1141                 if (err)
1142                         extent_io_tree_panic(tree, err);
1143                 cache_state(prealloc, cached_state);
1144                 prealloc = NULL;
1145                 goto out;
1146         }
1147         state = rb_entry(node, struct extent_state, rb_node);
1148 hit_next:
1149         last_start = state->start;
1150         last_end = state->end;
1151
1152         /*
1153          * | ---- desired range ---- |
1154          * | state |
1155          *
1156          * Just lock what we found and keep going
1157          */
1158         if (state->start == start && state->end <= end) {
1159                 set_state_bits(tree, state, &bits, NULL);
1160                 cache_state(state, cached_state);
1161                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1162                 if (last_end == (u64)-1)
1163                         goto out;
1164                 start = last_end + 1;
1165                 if (start < end && state && state->start == start &&
1166                     !need_resched())
1167                         goto hit_next;
1168                 goto search_again;
1169         }
1170
1171         /*
1172          *     | ---- desired range ---- |
1173          * | state |
1174          *   or
1175          * | ------------- state -------------- |
1176          *
1177          * We need to split the extent we found, and may flip bits on
1178          * second half.
1179          *
1180          * If the extent we found extends past our
1181          * range, we just split and search again.  It'll get split
1182          * again the next time though.
1183          *
1184          * If the extent we found is inside our range, we set the
1185          * desired bit on it.
1186          */
1187         if (state->start < start) {
1188                 prealloc = alloc_extent_state_atomic(prealloc);
1189                 if (!prealloc) {
1190                         err = -ENOMEM;
1191                         goto out;
1192                 }
1193                 err = split_state(tree, state, prealloc, start);
1194                 if (err)
1195                         extent_io_tree_panic(tree, err);
1196                 prealloc = NULL;
1197                 if (err)
1198                         goto out;
1199                 if (state->end <= end) {
1200                         set_state_bits(tree, state, &bits, NULL);
1201                         cache_state(state, cached_state);
1202                         state = clear_state_bit(tree, state, &clear_bits, 0,
1203                                                 NULL);
1204                         if (last_end == (u64)-1)
1205                                 goto out;
1206                         start = last_end + 1;
1207                         if (start < end && state && state->start == start &&
1208                             !need_resched())
1209                                 goto hit_next;
1210                 }
1211                 goto search_again;
1212         }
1213         /*
1214          * | ---- desired range ---- |
1215          *     | state | or               | state |
1216          *
1217          * There's a hole, we need to insert something in it and
1218          * ignore the extent we found.
1219          */
1220         if (state->start > start) {
1221                 u64 this_end;
1222                 if (end < last_start)
1223                         this_end = end;
1224                 else
1225                         this_end = last_start - 1;
1226
1227                 prealloc = alloc_extent_state_atomic(prealloc);
1228                 if (!prealloc) {
1229                         err = -ENOMEM;
1230                         goto out;
1231                 }
1232
1233                 /*
1234                  * Avoid to free 'prealloc' if it can be merged with
1235                  * the later extent.
1236                  */
1237                 err = insert_state(tree, prealloc, start, this_end,
1238                                    NULL, NULL, &bits, NULL);
1239                 if (err)
1240                         extent_io_tree_panic(tree, err);
1241                 cache_state(prealloc, cached_state);
1242                 prealloc = NULL;
1243                 start = this_end + 1;
1244                 goto search_again;
1245         }
1246         /*
1247          * | ---- desired range ---- |
1248          *                        | state |
1249          * We need to split the extent, and set the bit
1250          * on the first half
1251          */
1252         if (state->start <= end && state->end > end) {
1253                 prealloc = alloc_extent_state_atomic(prealloc);
1254                 if (!prealloc) {
1255                         err = -ENOMEM;
1256                         goto out;
1257                 }
1258
1259                 err = split_state(tree, state, prealloc, end + 1);
1260                 if (err)
1261                         extent_io_tree_panic(tree, err);
1262
1263                 set_state_bits(tree, prealloc, &bits, NULL);
1264                 cache_state(prealloc, cached_state);
1265                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1266                 prealloc = NULL;
1267                 goto out;
1268         }
1269
1270 search_again:
1271         if (start > end)
1272                 goto out;
1273         spin_unlock(&tree->lock);
1274         cond_resched();
1275         first_iteration = false;
1276         goto again;
1277
1278 out:
1279         spin_unlock(&tree->lock);
1280         if (prealloc)
1281                 free_extent_state(prealloc);
1282
1283         return err;
1284 }
1285
1286 /* wrappers around set/clear extent bit */
1287 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1288                            unsigned bits, 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, GFP_NOFS,
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, struct extent_changeset *changeset)
1312 {
1313         /*
1314          * Don't support EXTENT_LOCKED case, same reason as
1315          * set_record_extent_bits().
1316          */
1317         BUG_ON(bits & EXTENT_LOCKED);
1318
1319         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1320                                   changeset);
1321 }
1322
1323 /*
1324  * either insert or lock state struct between start and end use mask to tell
1325  * us if waiting is desired.
1326  */
1327 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1328                      struct extent_state **cached_state)
1329 {
1330         int err;
1331         u64 failed_start;
1332
1333         while (1) {
1334                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1335                                        EXTENT_LOCKED, &failed_start,
1336                                        cached_state, GFP_NOFS, NULL);
1337                 if (err == -EEXIST) {
1338                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1339                         start = failed_start;
1340                 } else
1341                         break;
1342                 WARN_ON(start > end);
1343         }
1344         return err;
1345 }
1346
1347 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1348 {
1349         int err;
1350         u64 failed_start;
1351
1352         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1353                                &failed_start, NULL, GFP_NOFS, NULL);
1354         if (err == -EEXIST) {
1355                 if (failed_start > start)
1356                         clear_extent_bit(tree, start, failed_start - 1,
1357                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1358                 return 0;
1359         }
1360         return 1;
1361 }
1362
1363 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1364 {
1365         unsigned long index = start >> PAGE_SHIFT;
1366         unsigned long end_index = end >> PAGE_SHIFT;
1367         struct page *page;
1368
1369         while (index <= end_index) {
1370                 page = find_get_page(inode->i_mapping, index);
1371                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1372                 clear_page_dirty_for_io(page);
1373                 put_page(page);
1374                 index++;
1375         }
1376 }
1377
1378 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1379 {
1380         unsigned long index = start >> PAGE_SHIFT;
1381         unsigned long end_index = end >> PAGE_SHIFT;
1382         struct page *page;
1383
1384         while (index <= end_index) {
1385                 page = find_get_page(inode->i_mapping, index);
1386                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1387                 __set_page_dirty_nobuffers(page);
1388                 account_page_redirty(page);
1389                 put_page(page);
1390                 index++;
1391         }
1392 }
1393
1394 /*
1395  * helper function to set both pages and extents in the tree writeback
1396  */
1397 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1398 {
1399         unsigned long index = start >> PAGE_SHIFT;
1400         unsigned long end_index = end >> PAGE_SHIFT;
1401         struct page *page;
1402
1403         while (index <= end_index) {
1404                 page = find_get_page(tree->mapping, index);
1405                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1406                 set_page_writeback(page);
1407                 put_page(page);
1408                 index++;
1409         }
1410 }
1411
1412 /* find the first state struct with 'bits' set after 'start', and
1413  * return it.  tree->lock must be held.  NULL will returned if
1414  * nothing was found after 'start'
1415  */
1416 static struct extent_state *
1417 find_first_extent_bit_state(struct extent_io_tree *tree,
1418                             u64 start, unsigned bits)
1419 {
1420         struct rb_node *node;
1421         struct extent_state *state;
1422
1423         /*
1424          * this search will find all the extents that end after
1425          * our range starts.
1426          */
1427         node = tree_search(tree, start);
1428         if (!node)
1429                 goto out;
1430
1431         while (1) {
1432                 state = rb_entry(node, struct extent_state, rb_node);
1433                 if (state->end >= start && (state->state & bits))
1434                         return state;
1435
1436                 node = rb_next(node);
1437                 if (!node)
1438                         break;
1439         }
1440 out:
1441         return NULL;
1442 }
1443
1444 /*
1445  * find the first offset in the io tree with 'bits' set. zero is
1446  * returned if we find something, and *start_ret and *end_ret are
1447  * set to reflect the state struct that was found.
1448  *
1449  * If nothing was found, 1 is returned. If found something, return 0.
1450  */
1451 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1452                           u64 *start_ret, u64 *end_ret, unsigned bits,
1453                           struct extent_state **cached_state)
1454 {
1455         struct extent_state *state;
1456         struct rb_node *n;
1457         int ret = 1;
1458
1459         spin_lock(&tree->lock);
1460         if (cached_state && *cached_state) {
1461                 state = *cached_state;
1462                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1463                         n = rb_next(&state->rb_node);
1464                         while (n) {
1465                                 state = rb_entry(n, struct extent_state,
1466                                                  rb_node);
1467                                 if (state->state & bits)
1468                                         goto got_it;
1469                                 n = rb_next(n);
1470                         }
1471                         free_extent_state(*cached_state);
1472                         *cached_state = NULL;
1473                         goto out;
1474                 }
1475                 free_extent_state(*cached_state);
1476                 *cached_state = NULL;
1477         }
1478
1479         state = find_first_extent_bit_state(tree, start, bits);
1480 got_it:
1481         if (state) {
1482                 cache_state_if_flags(state, cached_state, 0);
1483                 *start_ret = state->start;
1484                 *end_ret = state->end;
1485                 ret = 0;
1486         }
1487 out:
1488         spin_unlock(&tree->lock);
1489         return ret;
1490 }
1491
1492 /*
1493  * find a contiguous range of bytes in the file marked as delalloc, not
1494  * more than 'max_bytes'.  start and end are used to return the range,
1495  *
1496  * 1 is returned if we find something, 0 if nothing was in the tree
1497  */
1498 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1499                                         u64 *start, u64 *end, u64 max_bytes,
1500                                         struct extent_state **cached_state)
1501 {
1502         struct rb_node *node;
1503         struct extent_state *state;
1504         u64 cur_start = *start;
1505         u64 found = 0;
1506         u64 total_bytes = 0;
1507
1508         spin_lock(&tree->lock);
1509
1510         /*
1511          * this search will find all the extents that end after
1512          * our range starts.
1513          */
1514         node = tree_search(tree, cur_start);
1515         if (!node) {
1516                 if (!found)
1517                         *end = (u64)-1;
1518                 goto out;
1519         }
1520
1521         while (1) {
1522                 state = rb_entry(node, struct extent_state, rb_node);
1523                 if (found && (state->start != cur_start ||
1524                               (state->state & EXTENT_BOUNDARY))) {
1525                         goto out;
1526                 }
1527                 if (!(state->state & EXTENT_DELALLOC)) {
1528                         if (!found)
1529                                 *end = state->end;
1530                         goto out;
1531                 }
1532                 if (!found) {
1533                         *start = state->start;
1534                         *cached_state = state;
1535                         atomic_inc(&state->refs);
1536                 }
1537                 found++;
1538                 *end = state->end;
1539                 cur_start = state->end + 1;
1540                 node = rb_next(node);
1541                 total_bytes += state->end - state->start + 1;
1542                 if (total_bytes >= max_bytes)
1543                         break;
1544                 if (!node)
1545                         break;
1546         }
1547 out:
1548         spin_unlock(&tree->lock);
1549         return found;
1550 }
1551
1552 static noinline void __unlock_for_delalloc(struct inode *inode,
1553                                            struct page *locked_page,
1554                                            u64 start, u64 end)
1555 {
1556         int ret;
1557         struct page *pages[16];
1558         unsigned long index = start >> PAGE_SHIFT;
1559         unsigned long end_index = end >> PAGE_SHIFT;
1560         unsigned long nr_pages = end_index - index + 1;
1561         int i;
1562
1563         if (index == locked_page->index && end_index == index)
1564                 return;
1565
1566         while (nr_pages > 0) {
1567                 ret = find_get_pages_contig(inode->i_mapping, index,
1568                                      min_t(unsigned long, nr_pages,
1569                                      ARRAY_SIZE(pages)), pages);
1570                 for (i = 0; i < ret; i++) {
1571                         if (pages[i] != locked_page)
1572                                 unlock_page(pages[i]);
1573                         put_page(pages[i]);
1574                 }
1575                 nr_pages -= ret;
1576                 index += ret;
1577                 cond_resched();
1578         }
1579 }
1580
1581 static noinline int lock_delalloc_pages(struct inode *inode,
1582                                         struct page *locked_page,
1583                                         u64 delalloc_start,
1584                                         u64 delalloc_end)
1585 {
1586         unsigned long index = delalloc_start >> PAGE_SHIFT;
1587         unsigned long start_index = index;
1588         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1589         unsigned long pages_locked = 0;
1590         struct page *pages[16];
1591         unsigned long nrpages;
1592         int ret;
1593         int i;
1594
1595         /* the caller is responsible for locking the start index */
1596         if (index == locked_page->index && index == end_index)
1597                 return 0;
1598
1599         /* skip the page at the start index */
1600         nrpages = end_index - index + 1;
1601         while (nrpages > 0) {
1602                 ret = find_get_pages_contig(inode->i_mapping, index,
1603                                      min_t(unsigned long,
1604                                      nrpages, ARRAY_SIZE(pages)), pages);
1605                 if (ret == 0) {
1606                         ret = -EAGAIN;
1607                         goto done;
1608                 }
1609                 /* now we have an array of pages, lock them all */
1610                 for (i = 0; i < ret; i++) {
1611                         /*
1612                          * the caller is taking responsibility for
1613                          * locked_page
1614                          */
1615                         if (pages[i] != locked_page) {
1616                                 lock_page(pages[i]);
1617                                 if (!PageDirty(pages[i]) ||
1618                                     pages[i]->mapping != inode->i_mapping) {
1619                                         ret = -EAGAIN;
1620                                         unlock_page(pages[i]);
1621                                         put_page(pages[i]);
1622                                         goto done;
1623                                 }
1624                         }
1625                         put_page(pages[i]);
1626                         pages_locked++;
1627                 }
1628                 nrpages -= ret;
1629                 index += ret;
1630                 cond_resched();
1631         }
1632         ret = 0;
1633 done:
1634         if (ret && pages_locked) {
1635                 __unlock_for_delalloc(inode, locked_page,
1636                               delalloc_start,
1637                               ((u64)(start_index + pages_locked - 1)) <<
1638                               PAGE_SHIFT);
1639         }
1640         return ret;
1641 }
1642
1643 /*
1644  * find a contiguous range of bytes in the file marked as delalloc, not
1645  * more than 'max_bytes'.  start and end are used to return the range,
1646  *
1647  * 1 is returned if we find something, 0 if nothing was in the tree
1648  */
1649 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1650                                     struct extent_io_tree *tree,
1651                                     struct page *locked_page, u64 *start,
1652                                     u64 *end, u64 max_bytes)
1653 {
1654         u64 delalloc_start;
1655         u64 delalloc_end;
1656         u64 found;
1657         struct extent_state *cached_state = NULL;
1658         int ret;
1659         int loops = 0;
1660
1661 again:
1662         /* step one, find a bunch of delalloc bytes starting at start */
1663         delalloc_start = *start;
1664         delalloc_end = 0;
1665         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1666                                     max_bytes, &cached_state);
1667         if (!found || delalloc_end <= *start) {
1668                 *start = delalloc_start;
1669                 *end = delalloc_end;
1670                 free_extent_state(cached_state);
1671                 return 0;
1672         }
1673
1674         /*
1675          * start comes from the offset of locked_page.  We have to lock
1676          * pages in order, so we can't process delalloc bytes before
1677          * locked_page
1678          */
1679         if (delalloc_start < *start)
1680                 delalloc_start = *start;
1681
1682         /*
1683          * make sure to limit the number of pages we try to lock down
1684          */
1685         if (delalloc_end + 1 - delalloc_start > max_bytes)
1686                 delalloc_end = delalloc_start + max_bytes - 1;
1687
1688         /* step two, lock all the pages after the page that has start */
1689         ret = lock_delalloc_pages(inode, locked_page,
1690                                   delalloc_start, delalloc_end);
1691         if (ret == -EAGAIN) {
1692                 /* some of the pages are gone, lets avoid looping by
1693                  * shortening the size of the delalloc range we're searching
1694                  */
1695                 free_extent_state(cached_state);
1696                 cached_state = NULL;
1697                 if (!loops) {
1698                         max_bytes = PAGE_SIZE;
1699                         loops = 1;
1700                         goto again;
1701                 } else {
1702                         found = 0;
1703                         goto out_failed;
1704                 }
1705         }
1706         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1707
1708         /* step three, lock the state bits for the whole range */
1709         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1710
1711         /* then test to make sure it is all still delalloc */
1712         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1713                              EXTENT_DELALLOC, 1, cached_state);
1714         if (!ret) {
1715                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1716                                      &cached_state, GFP_NOFS);
1717                 __unlock_for_delalloc(inode, locked_page,
1718                               delalloc_start, delalloc_end);
1719                 cond_resched();
1720                 goto again;
1721         }
1722         free_extent_state(cached_state);
1723         *start = delalloc_start;
1724         *end = delalloc_end;
1725 out_failed:
1726         return found;
1727 }
1728
1729 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1730                                  u64 delalloc_end, struct page *locked_page,
1731                                  unsigned clear_bits,
1732                                  unsigned long page_ops)
1733 {
1734         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1735         int ret;
1736         struct page *pages[16];
1737         unsigned long index = start >> PAGE_SHIFT;
1738         unsigned long end_index = end >> PAGE_SHIFT;
1739         unsigned long nr_pages = end_index - index + 1;
1740         int i;
1741
1742         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1743         if (page_ops == 0)
1744                 return;
1745
1746         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1747                 mapping_set_error(inode->i_mapping, -EIO);
1748
1749         while (nr_pages > 0) {
1750                 ret = find_get_pages_contig(inode->i_mapping, index,
1751                                      min_t(unsigned long,
1752                                      nr_pages, ARRAY_SIZE(pages)), pages);
1753                 for (i = 0; i < ret; i++) {
1754
1755                         if (page_ops & PAGE_SET_PRIVATE2)
1756                                 SetPagePrivate2(pages[i]);
1757
1758                         if (pages[i] == locked_page) {
1759                                 put_page(pages[i]);
1760                                 continue;
1761                         }
1762                         if (page_ops & PAGE_CLEAR_DIRTY)
1763                                 clear_page_dirty_for_io(pages[i]);
1764                         if (page_ops & PAGE_SET_WRITEBACK)
1765                                 set_page_writeback(pages[i]);
1766                         if (page_ops & PAGE_SET_ERROR)
1767                                 SetPageError(pages[i]);
1768                         if (page_ops & PAGE_END_WRITEBACK)
1769                                 end_page_writeback(pages[i]);
1770                         if (page_ops & PAGE_UNLOCK)
1771                                 unlock_page(pages[i]);
1772                         put_page(pages[i]);
1773                 }
1774                 nr_pages -= ret;
1775                 index += ret;
1776                 cond_resched();
1777         }
1778 }
1779
1780 /*
1781  * count the number of bytes in the tree that have a given bit(s)
1782  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1783  * cached.  The total number found is returned.
1784  */
1785 u64 count_range_bits(struct extent_io_tree *tree,
1786                      u64 *start, u64 search_end, u64 max_bytes,
1787                      unsigned bits, int contig)
1788 {
1789         struct rb_node *node;
1790         struct extent_state *state;
1791         u64 cur_start = *start;
1792         u64 total_bytes = 0;
1793         u64 last = 0;
1794         int found = 0;
1795
1796         if (WARN_ON(search_end <= cur_start))
1797                 return 0;
1798
1799         spin_lock(&tree->lock);
1800         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1801                 total_bytes = tree->dirty_bytes;
1802                 goto out;
1803         }
1804         /*
1805          * this search will find all the extents that end after
1806          * our range starts.
1807          */
1808         node = tree_search(tree, cur_start);
1809         if (!node)
1810                 goto out;
1811
1812         while (1) {
1813                 state = rb_entry(node, struct extent_state, rb_node);
1814                 if (state->start > search_end)
1815                         break;
1816                 if (contig && found && state->start > last + 1)
1817                         break;
1818                 if (state->end >= cur_start && (state->state & bits) == bits) {
1819                         total_bytes += min(search_end, state->end) + 1 -
1820                                        max(cur_start, state->start);
1821                         if (total_bytes >= max_bytes)
1822                                 break;
1823                         if (!found) {
1824                                 *start = max(cur_start, state->start);
1825                                 found = 1;
1826                         }
1827                         last = state->end;
1828                 } else if (contig && found) {
1829                         break;
1830                 }
1831                 node = rb_next(node);
1832                 if (!node)
1833                         break;
1834         }
1835 out:
1836         spin_unlock(&tree->lock);
1837         return total_bytes;
1838 }
1839
1840 /*
1841  * set the private field for a given byte offset in the tree.  If there isn't
1842  * an extent_state there already, this does nothing.
1843  */
1844 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1845                 struct io_failure_record *failrec)
1846 {
1847         struct rb_node *node;
1848         struct extent_state *state;
1849         int ret = 0;
1850
1851         spin_lock(&tree->lock);
1852         /*
1853          * this search will find all the extents that end after
1854          * our range starts.
1855          */
1856         node = tree_search(tree, start);
1857         if (!node) {
1858                 ret = -ENOENT;
1859                 goto out;
1860         }
1861         state = rb_entry(node, struct extent_state, rb_node);
1862         if (state->start != start) {
1863                 ret = -ENOENT;
1864                 goto out;
1865         }
1866         state->failrec = failrec;
1867 out:
1868         spin_unlock(&tree->lock);
1869         return ret;
1870 }
1871
1872 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1873                 struct io_failure_record **failrec)
1874 {
1875         struct rb_node *node;
1876         struct extent_state *state;
1877         int ret = 0;
1878
1879         spin_lock(&tree->lock);
1880         /*
1881          * this search will find all the extents that end after
1882          * our range starts.
1883          */
1884         node = tree_search(tree, start);
1885         if (!node) {
1886                 ret = -ENOENT;
1887                 goto out;
1888         }
1889         state = rb_entry(node, struct extent_state, rb_node);
1890         if (state->start != start) {
1891                 ret = -ENOENT;
1892                 goto out;
1893         }
1894         *failrec = state->failrec;
1895 out:
1896         spin_unlock(&tree->lock);
1897         return ret;
1898 }
1899
1900 /*
1901  * searches a range in the state tree for a given mask.
1902  * If 'filled' == 1, this returns 1 only if every extent in the tree
1903  * has the bits set.  Otherwise, 1 is returned if any bit in the
1904  * range is found set.
1905  */
1906 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1907                    unsigned bits, int filled, struct extent_state *cached)
1908 {
1909         struct extent_state *state = NULL;
1910         struct rb_node *node;
1911         int bitset = 0;
1912
1913         spin_lock(&tree->lock);
1914         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1915             cached->end > start)
1916                 node = &cached->rb_node;
1917         else
1918                 node = tree_search(tree, start);
1919         while (node && start <= end) {
1920                 state = rb_entry(node, struct extent_state, rb_node);
1921
1922                 if (filled && state->start > start) {
1923                         bitset = 0;
1924                         break;
1925                 }
1926
1927                 if (state->start > end)
1928                         break;
1929
1930                 if (state->state & bits) {
1931                         bitset = 1;
1932                         if (!filled)
1933                                 break;
1934                 } else if (filled) {
1935                         bitset = 0;
1936                         break;
1937                 }
1938
1939                 if (state->end == (u64)-1)
1940                         break;
1941
1942                 start = state->end + 1;
1943                 if (start > end)
1944                         break;
1945                 node = rb_next(node);
1946                 if (!node) {
1947                         if (filled)
1948                                 bitset = 0;
1949                         break;
1950                 }
1951         }
1952         spin_unlock(&tree->lock);
1953         return bitset;
1954 }
1955
1956 /*
1957  * helper function to set a given page up to date if all the
1958  * extents in the tree for that page are up to date
1959  */
1960 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1961 {
1962         u64 start = page_offset(page);
1963         u64 end = start + PAGE_SIZE - 1;
1964         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1965                 SetPageUptodate(page);
1966 }
1967
1968 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1969 {
1970         int ret;
1971         int err = 0;
1972         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1973
1974         set_state_failrec(failure_tree, rec->start, NULL);
1975         ret = clear_extent_bits(failure_tree, rec->start,
1976                                 rec->start + rec->len - 1,
1977                                 EXTENT_LOCKED | EXTENT_DIRTY);
1978         if (ret)
1979                 err = ret;
1980
1981         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1982                                 rec->start + rec->len - 1,
1983                                 EXTENT_DAMAGED);
1984         if (ret && !err)
1985                 err = ret;
1986
1987         kfree(rec);
1988         return err;
1989 }
1990
1991 /*
1992  * this bypasses the standard btrfs submit functions deliberately, as
1993  * the standard behavior is to write all copies in a raid setup. here we only
1994  * want to write the one bad copy. so we do the mapping for ourselves and issue
1995  * submit_bio directly.
1996  * to avoid any synchronization issues, wait for the data after writing, which
1997  * actually prevents the read that triggered the error from finishing.
1998  * currently, there can be no more than two copies of every data bit. thus,
1999  * exactly one rewrite is required.
2000  */
2001 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2002                       struct page *page, unsigned int pg_offset, int mirror_num)
2003 {
2004         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2005         struct bio *bio;
2006         struct btrfs_device *dev;
2007         u64 map_length = 0;
2008         u64 sector;
2009         struct btrfs_bio *bbio = NULL;
2010         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2011         int ret;
2012
2013         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2014         BUG_ON(!mirror_num);
2015
2016         /* we can't repair anything in raid56 yet */
2017         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2018                 return 0;
2019
2020         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2021         if (!bio)
2022                 return -EIO;
2023         bio->bi_iter.bi_size = 0;
2024         map_length = length;
2025
2026         /*
2027          * Avoid races with device replace and make sure our bbio has devices
2028          * associated to its stripes that don't go away while we are doing the
2029          * read repair operation.
2030          */
2031         btrfs_bio_counter_inc_blocked(fs_info);
2032         ret = btrfs_map_block(fs_info, WRITE, logical,
2033                               &map_length, &bbio, mirror_num);
2034         if (ret) {
2035                 btrfs_bio_counter_dec(fs_info);
2036                 bio_put(bio);
2037                 return -EIO;
2038         }
2039         BUG_ON(mirror_num != bbio->mirror_num);
2040         sector = bbio->stripes[mirror_num-1].physical >> 9;
2041         bio->bi_iter.bi_sector = sector;
2042         dev = bbio->stripes[mirror_num-1].dev;
2043         btrfs_put_bbio(bbio);
2044         if (!dev || !dev->bdev || !dev->writeable) {
2045                 btrfs_bio_counter_dec(fs_info);
2046                 bio_put(bio);
2047                 return -EIO;
2048         }
2049         bio->bi_bdev = dev->bdev;
2050         bio_set_op_attrs(bio, REQ_OP_WRITE, WRITE_SYNC);
2051         bio_add_page(bio, page, length, pg_offset);
2052
2053         if (btrfsic_submit_bio_wait(bio)) {
2054                 /* try to remap that extent elsewhere? */
2055                 btrfs_bio_counter_dec(fs_info);
2056                 bio_put(bio);
2057                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2058                 return -EIO;
2059         }
2060
2061         btrfs_info_rl_in_rcu(fs_info,
2062                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2063                                   btrfs_ino(inode), start,
2064                                   rcu_str_deref(dev->name), sector);
2065         btrfs_bio_counter_dec(fs_info);
2066         bio_put(bio);
2067         return 0;
2068 }
2069
2070 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2071                          int mirror_num)
2072 {
2073         u64 start = eb->start;
2074         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2075         int ret = 0;
2076
2077         if (root->fs_info->sb->s_flags & MS_RDONLY)
2078                 return -EROFS;
2079
2080         for (i = 0; i < num_pages; i++) {
2081                 struct page *p = eb->pages[i];
2082
2083                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2084                                         PAGE_SIZE, start, p,
2085                                         start - page_offset(p), mirror_num);
2086                 if (ret)
2087                         break;
2088                 start += PAGE_SIZE;
2089         }
2090
2091         return ret;
2092 }
2093
2094 /*
2095  * each time an IO finishes, we do a fast check in the IO failure tree
2096  * to see if we need to process or clean up an io_failure_record
2097  */
2098 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2099                      unsigned int pg_offset)
2100 {
2101         u64 private;
2102         struct io_failure_record *failrec;
2103         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2104         struct extent_state *state;
2105         int num_copies;
2106         int ret;
2107
2108         private = 0;
2109         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2110                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2111         if (!ret)
2112                 return 0;
2113
2114         ret = get_state_failrec(&BTRFS_I(inode)->io_failure_tree, start,
2115                         &failrec);
2116         if (ret)
2117                 return 0;
2118
2119         BUG_ON(!failrec->this_mirror);
2120
2121         if (failrec->in_validation) {
2122                 /* there was no real error, just free the record */
2123                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2124                          failrec->start);
2125                 goto out;
2126         }
2127         if (fs_info->sb->s_flags & MS_RDONLY)
2128                 goto out;
2129
2130         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2131         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2132                                             failrec->start,
2133                                             EXTENT_LOCKED);
2134         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2135
2136         if (state && state->start <= failrec->start &&
2137             state->end >= failrec->start + failrec->len - 1) {
2138                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2139                                               failrec->len);
2140                 if (num_copies > 1)  {
2141                         repair_io_failure(inode, start, failrec->len,
2142                                           failrec->logical, page,
2143                                           pg_offset, failrec->failed_mirror);
2144                 }
2145         }
2146
2147 out:
2148         free_io_failure(inode, failrec);
2149
2150         return 0;
2151 }
2152
2153 /*
2154  * Can be called when
2155  * - hold extent lock
2156  * - under ordered extent
2157  * - the inode is freeing
2158  */
2159 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2160 {
2161         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2162         struct io_failure_record *failrec;
2163         struct extent_state *state, *next;
2164
2165         if (RB_EMPTY_ROOT(&failure_tree->state))
2166                 return;
2167
2168         spin_lock(&failure_tree->lock);
2169         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2170         while (state) {
2171                 if (state->start > end)
2172                         break;
2173
2174                 ASSERT(state->end <= end);
2175
2176                 next = next_state(state);
2177
2178                 failrec = state->failrec;
2179                 free_extent_state(state);
2180                 kfree(failrec);
2181
2182                 state = next;
2183         }
2184         spin_unlock(&failure_tree->lock);
2185 }
2186
2187 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2188                 struct io_failure_record **failrec_ret)
2189 {
2190         struct io_failure_record *failrec;
2191         struct extent_map *em;
2192         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2193         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2194         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2195         int ret;
2196         u64 logical;
2197
2198         ret = get_state_failrec(failure_tree, start, &failrec);
2199         if (ret) {
2200                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2201                 if (!failrec)
2202                         return -ENOMEM;
2203
2204                 failrec->start = start;
2205                 failrec->len = end - start + 1;
2206                 failrec->this_mirror = 0;
2207                 failrec->bio_flags = 0;
2208                 failrec->in_validation = 0;
2209
2210                 read_lock(&em_tree->lock);
2211                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2212                 if (!em) {
2213                         read_unlock(&em_tree->lock);
2214                         kfree(failrec);
2215                         return -EIO;
2216                 }
2217
2218                 if (em->start > start || em->start + em->len <= start) {
2219                         free_extent_map(em);
2220                         em = NULL;
2221                 }
2222                 read_unlock(&em_tree->lock);
2223                 if (!em) {
2224                         kfree(failrec);
2225                         return -EIO;
2226                 }
2227
2228                 logical = start - em->start;
2229                 logical = em->block_start + logical;
2230                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2231                         logical = em->block_start;
2232                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2233                         extent_set_compress_type(&failrec->bio_flags,
2234                                                  em->compress_type);
2235                 }
2236
2237                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2238                          logical, start, failrec->len);
2239
2240                 failrec->logical = logical;
2241                 free_extent_map(em);
2242
2243                 /* set the bits in the private failure tree */
2244                 ret = set_extent_bits(failure_tree, start, end,
2245                                         EXTENT_LOCKED | EXTENT_DIRTY);
2246                 if (ret >= 0)
2247                         ret = set_state_failrec(failure_tree, start, failrec);
2248                 /* set the bits in the inode's tree */
2249                 if (ret >= 0)
2250                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2251                 if (ret < 0) {
2252                         kfree(failrec);
2253                         return ret;
2254                 }
2255         } else {
2256                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2257                          failrec->logical, failrec->start, failrec->len,
2258                          failrec->in_validation);
2259                 /*
2260                  * when data can be on disk more than twice, add to failrec here
2261                  * (e.g. with a list for failed_mirror) to make
2262                  * clean_io_failure() clean all those errors at once.
2263                  */
2264         }
2265
2266         *failrec_ret = failrec;
2267
2268         return 0;
2269 }
2270
2271 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2272                            struct io_failure_record *failrec, int failed_mirror)
2273 {
2274         int num_copies;
2275
2276         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2277                                       failrec->logical, failrec->len);
2278         if (num_copies == 1) {
2279                 /*
2280                  * we only have a single copy of the data, so don't bother with
2281                  * all the retry and error correction code that follows. no
2282                  * matter what the error is, it is very likely to persist.
2283                  */
2284                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2285                          num_copies, failrec->this_mirror, failed_mirror);
2286                 return 0;
2287         }
2288
2289         /*
2290          * there are two premises:
2291          *      a) deliver good data to the caller
2292          *      b) correct the bad sectors on disk
2293          */
2294         if (failed_bio->bi_vcnt > 1) {
2295                 /*
2296                  * to fulfill b), we need to know the exact failing sectors, as
2297                  * we don't want to rewrite any more than the failed ones. thus,
2298                  * we need separate read requests for the failed bio
2299                  *
2300                  * if the following BUG_ON triggers, our validation request got
2301                  * merged. we need separate requests for our algorithm to work.
2302                  */
2303                 BUG_ON(failrec->in_validation);
2304                 failrec->in_validation = 1;
2305                 failrec->this_mirror = failed_mirror;
2306         } else {
2307                 /*
2308                  * we're ready to fulfill a) and b) alongside. get a good copy
2309                  * of the failed sector and if we succeed, we have setup
2310                  * everything for repair_io_failure to do the rest for us.
2311                  */
2312                 if (failrec->in_validation) {
2313                         BUG_ON(failrec->this_mirror != failed_mirror);
2314                         failrec->in_validation = 0;
2315                         failrec->this_mirror = 0;
2316                 }
2317                 failrec->failed_mirror = failed_mirror;
2318                 failrec->this_mirror++;
2319                 if (failrec->this_mirror == failed_mirror)
2320                         failrec->this_mirror++;
2321         }
2322
2323         if (failrec->this_mirror > num_copies) {
2324                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2325                          num_copies, failrec->this_mirror, failed_mirror);
2326                 return 0;
2327         }
2328
2329         return 1;
2330 }
2331
2332
2333 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2334                                     struct io_failure_record *failrec,
2335                                     struct page *page, int pg_offset, int icsum,
2336                                     bio_end_io_t *endio_func, void *data)
2337 {
2338         struct bio *bio;
2339         struct btrfs_io_bio *btrfs_failed_bio;
2340         struct btrfs_io_bio *btrfs_bio;
2341
2342         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2343         if (!bio)
2344                 return NULL;
2345
2346         bio->bi_end_io = endio_func;
2347         bio->bi_iter.bi_sector = failrec->logical >> 9;
2348         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2349         bio->bi_iter.bi_size = 0;
2350         bio->bi_private = data;
2351
2352         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2353         if (btrfs_failed_bio->csum) {
2354                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2355                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2356
2357                 btrfs_bio = btrfs_io_bio(bio);
2358                 btrfs_bio->csum = btrfs_bio->csum_inline;
2359                 icsum *= csum_size;
2360                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2361                        csum_size);
2362         }
2363
2364         bio_add_page(bio, page, failrec->len, pg_offset);
2365
2366         return bio;
2367 }
2368
2369 /*
2370  * this is a generic handler for readpage errors (default
2371  * readpage_io_failed_hook). if other copies exist, read those and write back
2372  * good data to the failed position. does not investigate in remapping the
2373  * failed extent elsewhere, hoping the device will be smart enough to do this as
2374  * needed
2375  */
2376
2377 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2378                               struct page *page, u64 start, u64 end,
2379                               int failed_mirror)
2380 {
2381         struct io_failure_record *failrec;
2382         struct inode *inode = page->mapping->host;
2383         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2384         struct bio *bio;
2385         int read_mode;
2386         int ret;
2387
2388         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2389
2390         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2391         if (ret)
2392                 return ret;
2393
2394         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2395         if (!ret) {
2396                 free_io_failure(inode, failrec);
2397                 return -EIO;
2398         }
2399
2400         if (failed_bio->bi_vcnt > 1)
2401                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2402         else
2403                 read_mode = READ_SYNC;
2404
2405         phy_offset >>= inode->i_sb->s_blocksize_bits;
2406         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2407                                       start - page_offset(page),
2408                                       (int)phy_offset, failed_bio->bi_end_io,
2409                                       NULL);
2410         if (!bio) {
2411                 free_io_failure(inode, failrec);
2412                 return -EIO;
2413         }
2414         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2415
2416         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2417                  read_mode, failrec->this_mirror, failrec->in_validation);
2418
2419         ret = tree->ops->submit_bio_hook(inode, bio, failrec->this_mirror,
2420                                          failrec->bio_flags, 0);
2421         if (ret) {
2422                 free_io_failure(inode, failrec);
2423                 bio_put(bio);
2424         }
2425
2426         return ret;
2427 }
2428
2429 /* lots and lots of room for performance fixes in the end_bio funcs */
2430
2431 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2432 {
2433         int uptodate = (err == 0);
2434         struct extent_io_tree *tree;
2435         int ret = 0;
2436
2437         tree = &BTRFS_I(page->mapping->host)->io_tree;
2438
2439         if (tree->ops && tree->ops->writepage_end_io_hook) {
2440                 ret = tree->ops->writepage_end_io_hook(page, start,
2441                                                end, NULL, uptodate);
2442                 if (ret)
2443                         uptodate = 0;
2444         }
2445
2446         if (!uptodate) {
2447                 ClearPageUptodate(page);
2448                 SetPageError(page);
2449                 ret = ret < 0 ? ret : -EIO;
2450                 mapping_set_error(page->mapping, ret);
2451         }
2452 }
2453
2454 /*
2455  * after a writepage IO is done, we need to:
2456  * clear the uptodate bits on error
2457  * clear the writeback bits in the extent tree for this IO
2458  * end_page_writeback if the page has no more pending IO
2459  *
2460  * Scheduling is not allowed, so the extent state tree is expected
2461  * to have one and only one object corresponding to this IO.
2462  */
2463 static void end_bio_extent_writepage(struct bio *bio)
2464 {
2465         struct bio_vec *bvec;
2466         u64 start;
2467         u64 end;
2468         int i;
2469
2470         bio_for_each_segment_all(bvec, bio, i) {
2471                 struct page *page = bvec->bv_page;
2472
2473                 /* We always issue full-page reads, but if some block
2474                  * in a page fails to read, blk_update_request() will
2475                  * advance bv_offset and adjust bv_len to compensate.
2476                  * Print a warning for nonzero offsets, and an error
2477                  * if they don't add up to a full page.  */
2478                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2479                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2480                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2481                                    "partial page write in btrfs with offset %u and length %u",
2482                                         bvec->bv_offset, bvec->bv_len);
2483                         else
2484                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2485                                    "incomplete page write in btrfs with offset %u and length %u",
2486                                         bvec->bv_offset, bvec->bv_len);
2487                 }
2488
2489                 start = page_offset(page);
2490                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2491
2492                 end_extent_writepage(page, bio->bi_error, start, end);
2493                 end_page_writeback(page);
2494         }
2495
2496         bio_put(bio);
2497 }
2498
2499 static void
2500 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2501                               int uptodate)
2502 {
2503         struct extent_state *cached = NULL;
2504         u64 end = start + len - 1;
2505
2506         if (uptodate && tree->track_uptodate)
2507                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2508         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2509 }
2510
2511 /*
2512  * after a readpage IO is done, we need to:
2513  * clear the uptodate bits on error
2514  * set the uptodate bits if things worked
2515  * set the page up to date if all extents in the tree are uptodate
2516  * clear the lock bit in the extent tree
2517  * unlock the page if there are no other extents locked for it
2518  *
2519  * Scheduling is not allowed, so the extent state tree is expected
2520  * to have one and only one object corresponding to this IO.
2521  */
2522 static void end_bio_extent_readpage(struct bio *bio)
2523 {
2524         struct bio_vec *bvec;
2525         int uptodate = !bio->bi_error;
2526         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2527         struct extent_io_tree *tree;
2528         u64 offset = 0;
2529         u64 start;
2530         u64 end;
2531         u64 len;
2532         u64 extent_start = 0;
2533         u64 extent_len = 0;
2534         int mirror;
2535         int ret;
2536         int i;
2537
2538         bio_for_each_segment_all(bvec, bio, i) {
2539                 struct page *page = bvec->bv_page;
2540                 struct inode *inode = page->mapping->host;
2541
2542                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u\n",
2543                          (u64)bio->bi_iter.bi_sector,
2544                          bio->bi_error, io_bio->mirror_num);
2545                 tree = &BTRFS_I(inode)->io_tree;
2546
2547                 /* We always issue full-page reads, but if some block
2548                  * in a page fails to read, blk_update_request() will
2549                  * advance bv_offset and adjust bv_len to compensate.
2550                  * Print a warning for nonzero offsets, and an error
2551                  * if they don't add up to a full page.  */
2552                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2553                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2554                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2555                                    "partial page read in btrfs with offset %u and length %u",
2556                                         bvec->bv_offset, bvec->bv_len);
2557                         else
2558                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2559                                    "incomplete page read in btrfs with offset %u and length %u",
2560                                         bvec->bv_offset, bvec->bv_len);
2561                 }
2562
2563                 start = page_offset(page);
2564                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2565                 len = bvec->bv_len;
2566
2567                 mirror = io_bio->mirror_num;
2568                 if (likely(uptodate && tree->ops &&
2569                            tree->ops->readpage_end_io_hook)) {
2570                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2571                                                               page, start, end,
2572                                                               mirror);
2573                         if (ret)
2574                                 uptodate = 0;
2575                         else
2576                                 clean_io_failure(inode, start, page, 0);
2577                 }
2578
2579                 if (likely(uptodate))
2580                         goto readpage_ok;
2581
2582                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2583                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2584                         if (!ret && !bio->bi_error)
2585                                 uptodate = 1;
2586                 } else {
2587                         /*
2588                          * The generic bio_readpage_error handles errors the
2589                          * following way: If possible, new read requests are
2590                          * created and submitted and will end up in
2591                          * end_bio_extent_readpage as well (if we're lucky, not
2592                          * in the !uptodate case). In that case it returns 0 and
2593                          * we just go on with the next page in our bio. If it
2594                          * can't handle the error it will return -EIO and we
2595                          * remain responsible for that page.
2596                          */
2597                         ret = bio_readpage_error(bio, offset, page, start, end,
2598                                                  mirror);
2599                         if (ret == 0) {
2600                                 uptodate = !bio->bi_error;
2601                                 offset += len;
2602                                 continue;
2603                         }
2604                 }
2605 readpage_ok:
2606                 if (likely(uptodate)) {
2607                         loff_t i_size = i_size_read(inode);
2608                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2609                         unsigned off;
2610
2611                         /* Zero out the end if this page straddles i_size */
2612                         off = i_size & (PAGE_SIZE-1);
2613                         if (page->index == end_index && off)
2614                                 zero_user_segment(page, off, PAGE_SIZE);
2615                         SetPageUptodate(page);
2616                 } else {
2617                         ClearPageUptodate(page);
2618                         SetPageError(page);
2619                 }
2620                 unlock_page(page);
2621                 offset += len;
2622
2623                 if (unlikely(!uptodate)) {
2624                         if (extent_len) {
2625                                 endio_readpage_release_extent(tree,
2626                                                               extent_start,
2627                                                               extent_len, 1);
2628                                 extent_start = 0;
2629                                 extent_len = 0;
2630                         }
2631                         endio_readpage_release_extent(tree, start,
2632                                                       end - start + 1, 0);
2633                 } else if (!extent_len) {
2634                         extent_start = start;
2635                         extent_len = end + 1 - start;
2636                 } else if (extent_start + extent_len == start) {
2637                         extent_len += end + 1 - start;
2638                 } else {
2639                         endio_readpage_release_extent(tree, extent_start,
2640                                                       extent_len, uptodate);
2641                         extent_start = start;
2642                         extent_len = end + 1 - start;
2643                 }
2644         }
2645
2646         if (extent_len)
2647                 endio_readpage_release_extent(tree, extent_start, extent_len,
2648                                               uptodate);
2649         if (io_bio->end_io)
2650                 io_bio->end_io(io_bio, bio->bi_error);
2651         bio_put(bio);
2652 }
2653
2654 /*
2655  * this allocates from the btrfs_bioset.  We're returning a bio right now
2656  * but you can call btrfs_io_bio for the appropriate container_of magic
2657  */
2658 struct bio *
2659 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2660                 gfp_t gfp_flags)
2661 {
2662         struct btrfs_io_bio *btrfs_bio;
2663         struct bio *bio;
2664
2665         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2666
2667         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2668                 while (!bio && (nr_vecs /= 2)) {
2669                         bio = bio_alloc_bioset(gfp_flags,
2670                                                nr_vecs, btrfs_bioset);
2671                 }
2672         }
2673
2674         if (bio) {
2675                 bio->bi_bdev = bdev;
2676                 bio->bi_iter.bi_sector = first_sector;
2677                 btrfs_bio = btrfs_io_bio(bio);
2678                 btrfs_bio->csum = NULL;
2679                 btrfs_bio->csum_allocated = NULL;
2680                 btrfs_bio->end_io = NULL;
2681         }
2682         return bio;
2683 }
2684
2685 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2686 {
2687         struct btrfs_io_bio *btrfs_bio;
2688         struct bio *new;
2689
2690         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2691         if (new) {
2692                 btrfs_bio = btrfs_io_bio(new);
2693                 btrfs_bio->csum = NULL;
2694                 btrfs_bio->csum_allocated = NULL;
2695                 btrfs_bio->end_io = NULL;
2696         }
2697         return new;
2698 }
2699
2700 /* this also allocates from the btrfs_bioset */
2701 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2702 {
2703         struct btrfs_io_bio *btrfs_bio;
2704         struct bio *bio;
2705
2706         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2707         if (bio) {
2708                 btrfs_bio = btrfs_io_bio(bio);
2709                 btrfs_bio->csum = NULL;
2710                 btrfs_bio->csum_allocated = NULL;
2711                 btrfs_bio->end_io = NULL;
2712         }
2713         return bio;
2714 }
2715
2716
2717 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2718                                        unsigned long bio_flags)
2719 {
2720         int ret = 0;
2721         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2722         struct page *page = bvec->bv_page;
2723         struct extent_io_tree *tree = bio->bi_private;
2724         u64 start;
2725
2726         start = page_offset(page) + bvec->bv_offset;
2727
2728         bio->bi_private = NULL;
2729         bio_get(bio);
2730
2731         if (tree->ops && tree->ops->submit_bio_hook)
2732                 ret = tree->ops->submit_bio_hook(page->mapping->host, bio,
2733                                            mirror_num, bio_flags, start);
2734         else
2735                 btrfsic_submit_bio(bio);
2736
2737         bio_put(bio);
2738         return ret;
2739 }
2740
2741 static int merge_bio(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(page, offset, size, bio,
2748                                                 bio_flags);
2749         return ret;
2750
2751 }
2752
2753 static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree,
2754                               struct writeback_control *wbc,
2755                               struct page *page, sector_t sector,
2756                               size_t size, unsigned long offset,
2757                               struct block_device *bdev,
2758                               struct bio **bio_ret,
2759                               unsigned long max_pages,
2760                               bio_end_io_t end_io_func,
2761                               int mirror_num,
2762                               unsigned long prev_bio_flags,
2763                               unsigned long bio_flags,
2764                               bool force_bio_submit)
2765 {
2766         int ret = 0;
2767         struct bio *bio;
2768         int contig = 0;
2769         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2770         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2771
2772         if (bio_ret && *bio_ret) {
2773                 bio = *bio_ret;
2774                 if (old_compressed)
2775                         contig = bio->bi_iter.bi_sector == sector;
2776                 else
2777                         contig = bio_end_sector(bio) == sector;
2778
2779                 if (prev_bio_flags != bio_flags || !contig ||
2780                     force_bio_submit ||
2781                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2782                     bio_add_page(bio, page, page_size, offset) < page_size) {
2783                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2784                         if (ret < 0) {
2785                                 *bio_ret = NULL;
2786                                 return ret;
2787                         }
2788                         bio = NULL;
2789                 } else {
2790                         if (wbc)
2791                                 wbc_account_io(wbc, page, page_size);
2792                         return 0;
2793                 }
2794         }
2795
2796         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2797                         GFP_NOFS | __GFP_HIGH);
2798         if (!bio)
2799                 return -ENOMEM;
2800
2801         bio_add_page(bio, page, page_size, offset);
2802         bio->bi_end_io = end_io_func;
2803         bio->bi_private = tree;
2804         bio_set_op_attrs(bio, op, op_flags);
2805         if (wbc) {
2806                 wbc_init_bio(wbc, bio);
2807                 wbc_account_io(wbc, page, page_size);
2808         }
2809
2810         if (bio_ret)
2811                 *bio_ret = bio;
2812         else
2813                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2814
2815         return ret;
2816 }
2817
2818 static void attach_extent_buffer_page(struct extent_buffer *eb,
2819                                       struct page *page)
2820 {
2821         if (!PagePrivate(page)) {
2822                 SetPagePrivate(page);
2823                 get_page(page);
2824                 set_page_private(page, (unsigned long)eb);
2825         } else {
2826                 WARN_ON(page->private != (unsigned long)eb);
2827         }
2828 }
2829
2830 void set_page_extent_mapped(struct page *page)
2831 {
2832         if (!PagePrivate(page)) {
2833                 SetPagePrivate(page);
2834                 get_page(page);
2835                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2836         }
2837 }
2838
2839 static struct extent_map *
2840 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2841                  u64 start, u64 len, get_extent_t *get_extent,
2842                  struct extent_map **em_cached)
2843 {
2844         struct extent_map *em;
2845
2846         if (em_cached && *em_cached) {
2847                 em = *em_cached;
2848                 if (extent_map_in_tree(em) && start >= em->start &&
2849                     start < extent_map_end(em)) {
2850                         atomic_inc(&em->refs);
2851                         return em;
2852                 }
2853
2854                 free_extent_map(em);
2855                 *em_cached = NULL;
2856         }
2857
2858         em = get_extent(inode, page, pg_offset, start, len, 0);
2859         if (em_cached && !IS_ERR_OR_NULL(em)) {
2860                 BUG_ON(*em_cached);
2861                 atomic_inc(&em->refs);
2862                 *em_cached = em;
2863         }
2864         return em;
2865 }
2866 /*
2867  * basic readpage implementation.  Locked extent state structs are inserted
2868  * into the tree that are removed when the IO is done (by the end_io
2869  * handlers)
2870  * XXX JDM: This needs looking at to ensure proper page locking
2871  * return 0 on success, otherwise return error
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 read_flags,
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_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 = 0;
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_SHIFT) {
2913                 char *userpage;
2914                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2915
2916                 if (zero_offset) {
2917                         iosize = PAGE_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_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_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(REQ_OP_READ, read_flags, tree, NULL,
3062                                          page, 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                         goto out;
3075                 }
3076                 cur = cur + iosize;
3077                 pg_offset += iosize;
3078         }
3079 out:
3080         if (!nr) {
3081                 if (!PageError(page))
3082                         SetPageUptodate(page);
3083                 unlock_page(page);
3084         }
3085         return ret;
3086 }
3087
3088 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3089                                              struct page *pages[], int nr_pages,
3090                                              u64 start, u64 end,
3091                                              get_extent_t *get_extent,
3092                                              struct extent_map **em_cached,
3093                                              struct bio **bio, int mirror_num,
3094                                              unsigned long *bio_flags,
3095                                              u64 *prev_em_start)
3096 {
3097         struct inode *inode;
3098         struct btrfs_ordered_extent *ordered;
3099         int index;
3100
3101         inode = pages[0]->mapping->host;
3102         while (1) {
3103                 lock_extent(tree, start, end);
3104                 ordered = btrfs_lookup_ordered_range(inode, start,
3105                                                      end - start + 1);
3106                 if (!ordered)
3107                         break;
3108                 unlock_extent(tree, start, end);
3109                 btrfs_start_ordered_extent(inode, ordered, 1);
3110                 btrfs_put_ordered_extent(ordered);
3111         }
3112
3113         for (index = 0; index < nr_pages; index++) {
3114                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3115                               mirror_num, bio_flags, 0, prev_em_start);
3116                 put_page(pages[index]);
3117         }
3118 }
3119
3120 static void __extent_readpages(struct extent_io_tree *tree,
3121                                struct page *pages[],
3122                                int nr_pages, get_extent_t *get_extent,
3123                                struct extent_map **em_cached,
3124                                struct bio **bio, int mirror_num,
3125                                unsigned long *bio_flags,
3126                                u64 *prev_em_start)
3127 {
3128         u64 start = 0;
3129         u64 end = 0;
3130         u64 page_start;
3131         int index;
3132         int first_index = 0;
3133
3134         for (index = 0; index < nr_pages; index++) {
3135                 page_start = page_offset(pages[index]);
3136                 if (!end) {
3137                         start = page_start;
3138                         end = start + PAGE_SIZE - 1;
3139                         first_index = index;
3140                 } else if (end + 1 == page_start) {
3141                         end += PAGE_SIZE;
3142                 } else {
3143                         __do_contiguous_readpages(tree, &pages[first_index],
3144                                                   index - first_index, start,
3145                                                   end, get_extent, em_cached,
3146                                                   bio, mirror_num, bio_flags,
3147                                                   prev_em_start);
3148                         start = page_start;
3149                         end = start + PAGE_SIZE - 1;
3150                         first_index = index;
3151                 }
3152         }
3153
3154         if (end)
3155                 __do_contiguous_readpages(tree, &pages[first_index],
3156                                           index - first_index, start,
3157                                           end, get_extent, em_cached, bio,
3158                                           mirror_num, bio_flags,
3159                                           prev_em_start);
3160 }
3161
3162 static int __extent_read_full_page(struct extent_io_tree *tree,
3163                                    struct page *page,
3164                                    get_extent_t *get_extent,
3165                                    struct bio **bio, int mirror_num,
3166                                    unsigned long *bio_flags, int read_flags)
3167 {
3168         struct inode *inode = page->mapping->host;
3169         struct btrfs_ordered_extent *ordered;
3170         u64 start = page_offset(page);
3171         u64 end = start + PAGE_SIZE - 1;
3172         int ret;
3173
3174         while (1) {
3175                 lock_extent(tree, start, end);
3176                 ordered = btrfs_lookup_ordered_range(inode, start,
3177                                                 PAGE_SIZE);
3178                 if (!ordered)
3179                         break;
3180                 unlock_extent(tree, start, end);
3181                 btrfs_start_ordered_extent(inode, ordered, 1);
3182                 btrfs_put_ordered_extent(ordered);
3183         }
3184
3185         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3186                             bio_flags, read_flags, NULL);
3187         return ret;
3188 }
3189
3190 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3191                             get_extent_t *get_extent, int mirror_num)
3192 {
3193         struct bio *bio = NULL;
3194         unsigned long bio_flags = 0;
3195         int ret;
3196
3197         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3198                                       &bio_flags, 0);
3199         if (bio)
3200                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3201         return ret;
3202 }
3203
3204 static void update_nr_written(struct page *page, struct writeback_control *wbc,
3205                               unsigned long nr_written)
3206 {
3207         wbc->nr_to_write -= nr_written;
3208 }
3209
3210 /*
3211  * helper for __extent_writepage, doing all of the delayed allocation setup.
3212  *
3213  * This returns 1 if our fill_delalloc function did all the work required
3214  * to write the page (copy into inline extent).  In this case the IO has
3215  * been started and the page is already unlocked.
3216  *
3217  * This returns 0 if all went well (page still locked)
3218  * This returns < 0 if there were errors (page still locked)
3219  */
3220 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3221                               struct page *page, struct writeback_control *wbc,
3222                               struct extent_page_data *epd,
3223                               u64 delalloc_start,
3224                               unsigned long *nr_written)
3225 {
3226         struct extent_io_tree *tree = epd->tree;
3227         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3228         u64 nr_delalloc;
3229         u64 delalloc_to_write = 0;
3230         u64 delalloc_end = 0;
3231         int ret;
3232         int page_started = 0;
3233
3234         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3235                 return 0;
3236
3237         while (delalloc_end < page_end) {
3238                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3239                                                page,
3240                                                &delalloc_start,
3241                                                &delalloc_end,
3242                                                BTRFS_MAX_EXTENT_SIZE);
3243                 if (nr_delalloc == 0) {
3244                         delalloc_start = delalloc_end + 1;
3245                         continue;
3246                 }
3247                 ret = tree->ops->fill_delalloc(inode, page,
3248                                                delalloc_start,
3249                                                delalloc_end,
3250                                                &page_started,
3251                                                nr_written);
3252                 /* File system has been set read-only */
3253                 if (ret) {
3254                         SetPageError(page);
3255                         /* fill_delalloc should be return < 0 for error
3256                          * but just in case, we use > 0 here meaning the
3257                          * IO is started, so we don't want to return > 0
3258                          * unless things are going well.
3259                          */
3260                         ret = ret < 0 ? ret : -EIO;
3261                         goto done;
3262                 }
3263                 /*
3264                  * delalloc_end is already one less than the total length, so
3265                  * we don't subtract one from PAGE_SIZE
3266                  */
3267                 delalloc_to_write += (delalloc_end - delalloc_start +
3268                                       PAGE_SIZE) >> PAGE_SHIFT;
3269                 delalloc_start = delalloc_end + 1;
3270         }
3271         if (wbc->nr_to_write < delalloc_to_write) {
3272                 int thresh = 8192;
3273
3274                 if (delalloc_to_write < thresh * 2)
3275                         thresh = delalloc_to_write;
3276                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3277                                          thresh);
3278         }
3279
3280         /* did the fill delalloc function already unlock and start
3281          * the IO?
3282          */
3283         if (page_started) {
3284                 /*
3285                  * we've unlocked the page, so we can't update
3286                  * the mapping's writeback index, just update
3287                  * nr_to_write.
3288                  */
3289                 wbc->nr_to_write -= *nr_written;
3290                 return 1;
3291         }
3292
3293         ret = 0;
3294
3295 done:
3296         return ret;
3297 }
3298
3299 /*
3300  * helper for __extent_writepage.  This calls the writepage start hooks,
3301  * and does the loop to map the page into extents and bios.
3302  *
3303  * We return 1 if the IO is started and the page is unlocked,
3304  * 0 if all went well (page still locked)
3305  * < 0 if there were errors (page still locked)
3306  */
3307 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3308                                  struct page *page,
3309                                  struct writeback_control *wbc,
3310                                  struct extent_page_data *epd,
3311                                  loff_t i_size,
3312                                  unsigned long nr_written,
3313                                  int write_flags, int *nr_ret)
3314 {
3315         struct extent_io_tree *tree = epd->tree;
3316         u64 start = page_offset(page);
3317         u64 page_end = start + PAGE_SIZE - 1;
3318         u64 end;
3319         u64 cur = start;
3320         u64 extent_offset;
3321         u64 block_start;
3322         u64 iosize;
3323         sector_t sector;
3324         struct extent_state *cached_state = NULL;
3325         struct extent_map *em;
3326         struct block_device *bdev;
3327         size_t pg_offset = 0;
3328         size_t blocksize;
3329         int ret = 0;
3330         int nr = 0;
3331         bool compressed;
3332
3333         if (tree->ops && tree->ops->writepage_start_hook) {
3334                 ret = tree->ops->writepage_start_hook(page, start,
3335                                                       page_end);
3336                 if (ret) {
3337                         /* Fixup worker will requeue */
3338                         if (ret == -EBUSY)
3339                                 wbc->pages_skipped++;
3340                         else
3341                                 redirty_page_for_writepage(wbc, page);
3342
3343                         update_nr_written(page, wbc, nr_written);
3344                         unlock_page(page);
3345                         ret = 1;
3346                         goto done_unlocked;
3347                 }
3348         }
3349
3350         /*
3351          * we don't want to touch the inode after unlocking the page,
3352          * so we update the mapping writeback index now
3353          */
3354         update_nr_written(page, wbc, nr_written + 1);
3355
3356         end = page_end;
3357         if (i_size <= start) {
3358                 if (tree->ops && tree->ops->writepage_end_io_hook)
3359                         tree->ops->writepage_end_io_hook(page, start,
3360                                                          page_end, NULL, 1);
3361                 goto done;
3362         }
3363
3364         blocksize = inode->i_sb->s_blocksize;
3365
3366         while (cur <= end) {
3367                 u64 em_end;
3368                 unsigned long max_nr;
3369
3370                 if (cur >= i_size) {
3371                         if (tree->ops && tree->ops->writepage_end_io_hook)
3372                                 tree->ops->writepage_end_io_hook(page, cur,
3373                                                          page_end, NULL, 1);
3374                         break;
3375                 }
3376                 em = epd->get_extent(inode, page, pg_offset, cur,
3377                                      end - cur + 1, 1);
3378                 if (IS_ERR_OR_NULL(em)) {
3379                         SetPageError(page);
3380                         ret = PTR_ERR_OR_ZERO(em);
3381                         break;
3382                 }
3383
3384                 extent_offset = cur - em->start;
3385                 em_end = extent_map_end(em);
3386                 BUG_ON(em_end <= cur);
3387                 BUG_ON(end < cur);
3388                 iosize = min(em_end - cur, end - cur + 1);
3389                 iosize = ALIGN(iosize, blocksize);
3390                 sector = (em->block_start + extent_offset) >> 9;
3391                 bdev = em->bdev;
3392                 block_start = em->block_start;
3393                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3394                 free_extent_map(em);
3395                 em = NULL;
3396
3397                 /*
3398                  * compressed and inline extents are written through other
3399                  * paths in the FS
3400                  */
3401                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3402                     block_start == EXTENT_MAP_INLINE) {
3403                         /*
3404                          * end_io notification does not happen here for
3405                          * compressed extents
3406                          */
3407                         if (!compressed && tree->ops &&
3408                             tree->ops->writepage_end_io_hook)
3409                                 tree->ops->writepage_end_io_hook(page, cur,
3410                                                          cur + iosize - 1,
3411                                                          NULL, 1);
3412                         else if (compressed) {
3413                                 /* we don't want to end_page_writeback on
3414                                  * a compressed extent.  this happens
3415                                  * elsewhere
3416                                  */
3417                                 nr++;
3418                         }
3419
3420                         cur += iosize;
3421                         pg_offset += iosize;
3422                         continue;
3423                 }
3424
3425                 max_nr = (i_size >> PAGE_SHIFT) + 1;
3426
3427                 set_range_writeback(tree, cur, cur + iosize - 1);
3428                 if (!PageWriteback(page)) {
3429                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3430                                    "page %lu not writeback, cur %llu end %llu",
3431                                page->index, cur, end);
3432                 }
3433
3434                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3435                                          page, sector, iosize, pg_offset,
3436                                          bdev, &epd->bio, max_nr,
3437                                          end_bio_extent_writepage,
3438                                          0, 0, 0, false);
3439                 if (ret)
3440                         SetPageError(page);
3441
3442                 cur = cur + iosize;
3443                 pg_offset += iosize;
3444                 nr++;
3445         }
3446 done:
3447         *nr_ret = nr;
3448
3449 done_unlocked:
3450
3451         /* drop our reference on any cached states */
3452         free_extent_state(cached_state);
3453         return ret;
3454 }
3455
3456 /*
3457  * the writepage semantics are similar to regular writepage.  extent
3458  * records are inserted to lock ranges in the tree, and as dirty areas
3459  * are found, they are marked writeback.  Then the lock bits are removed
3460  * and the end_io handler clears the writeback ranges
3461  */
3462 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3463                               void *data)
3464 {
3465         struct inode *inode = page->mapping->host;
3466         struct extent_page_data *epd = data;
3467         u64 start = page_offset(page);
3468         u64 page_end = start + PAGE_SIZE - 1;
3469         int ret;
3470         int nr = 0;
3471         size_t pg_offset = 0;
3472         loff_t i_size = i_size_read(inode);
3473         unsigned long end_index = i_size >> PAGE_SHIFT;
3474         int write_flags = 0;
3475         unsigned long nr_written = 0;
3476
3477         if (wbc->sync_mode == WB_SYNC_ALL)
3478                 write_flags = WRITE_SYNC;
3479
3480         trace___extent_writepage(page, inode, wbc);
3481
3482         WARN_ON(!PageLocked(page));
3483
3484         ClearPageError(page);
3485
3486         pg_offset = i_size & (PAGE_SIZE - 1);
3487         if (page->index > end_index ||
3488            (page->index == end_index && !pg_offset)) {
3489                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3490                 unlock_page(page);
3491                 return 0;
3492         }
3493
3494         if (page->index == end_index) {
3495                 char *userpage;
3496
3497                 userpage = kmap_atomic(page);
3498                 memset(userpage + pg_offset, 0,
3499                        PAGE_SIZE - pg_offset);
3500                 kunmap_atomic(userpage);
3501                 flush_dcache_page(page);
3502         }
3503
3504         pg_offset = 0;
3505
3506         set_page_extent_mapped(page);
3507
3508         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3509         if (ret == 1)
3510                 goto done_unlocked;
3511         if (ret)
3512                 goto done;
3513
3514         ret = __extent_writepage_io(inode, page, wbc, epd,
3515                                     i_size, nr_written, write_flags, &nr);
3516         if (ret == 1)
3517                 goto done_unlocked;
3518
3519 done:
3520         if (nr == 0) {
3521                 /* make sure the mapping tag for page dirty gets cleared */
3522                 set_page_writeback(page);
3523                 end_page_writeback(page);
3524         }
3525         if (PageError(page)) {
3526                 ret = ret < 0 ? ret : -EIO;
3527                 end_extent_writepage(page, ret, start, page_end);
3528         }
3529         unlock_page(page);
3530         return ret;
3531
3532 done_unlocked:
3533         return 0;
3534 }
3535
3536 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3537 {
3538         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3539                        TASK_UNINTERRUPTIBLE);
3540 }
3541
3542 static noinline_for_stack int
3543 lock_extent_buffer_for_io(struct extent_buffer *eb,
3544                           struct btrfs_fs_info *fs_info,
3545                           struct extent_page_data *epd)
3546 {
3547         unsigned long i, num_pages;
3548         int flush = 0;
3549         int ret = 0;
3550
3551         if (!btrfs_try_tree_write_lock(eb)) {
3552                 flush = 1;
3553                 flush_write_bio(epd);
3554                 btrfs_tree_lock(eb);
3555         }
3556
3557         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3558                 btrfs_tree_unlock(eb);
3559                 if (!epd->sync_io)
3560                         return 0;
3561                 if (!flush) {
3562                         flush_write_bio(epd);
3563                         flush = 1;
3564                 }
3565                 while (1) {
3566                         wait_on_extent_buffer_writeback(eb);
3567                         btrfs_tree_lock(eb);
3568                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3569                                 break;
3570                         btrfs_tree_unlock(eb);
3571                 }
3572         }
3573
3574         /*
3575          * We need to do this to prevent races in people who check if the eb is
3576          * under IO since we can end up having no IO bits set for a short period
3577          * of time.
3578          */
3579         spin_lock(&eb->refs_lock);
3580         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3581                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3582                 spin_unlock(&eb->refs_lock);
3583                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3584                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3585                                      -eb->len,
3586                                      fs_info->dirty_metadata_batch);
3587                 ret = 1;
3588         } else {
3589                 spin_unlock(&eb->refs_lock);
3590         }
3591
3592         btrfs_tree_unlock(eb);
3593
3594         if (!ret)
3595                 return ret;
3596
3597         num_pages = num_extent_pages(eb->start, eb->len);
3598         for (i = 0; i < num_pages; i++) {
3599                 struct page *p = eb->pages[i];
3600
3601                 if (!trylock_page(p)) {
3602                         if (!flush) {
3603                                 flush_write_bio(epd);
3604                                 flush = 1;
3605                         }
3606                         lock_page(p);
3607                 }
3608         }
3609
3610         return ret;
3611 }
3612
3613 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3614 {
3615         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3616         smp_mb__after_atomic();
3617         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3618 }
3619
3620 static void set_btree_ioerr(struct page *page)
3621 {
3622         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3623
3624         SetPageError(page);
3625         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3626                 return;
3627
3628         /*
3629          * If writeback for a btree extent that doesn't belong to a log tree
3630          * failed, increment the counter transaction->eb_write_errors.
3631          * We do this because while the transaction is running and before it's
3632          * committing (when we call filemap_fdata[write|wait]_range against
3633          * the btree inode), we might have
3634          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3635          * returns an error or an error happens during writeback, when we're
3636          * committing the transaction we wouldn't know about it, since the pages
3637          * can be no longer dirty nor marked anymore for writeback (if a
3638          * subsequent modification to the extent buffer didn't happen before the
3639          * transaction commit), which makes filemap_fdata[write|wait]_range not
3640          * able to find the pages tagged with SetPageError at transaction
3641          * commit time. So if this happens we must abort the transaction,
3642          * otherwise we commit a super block with btree roots that point to
3643          * btree nodes/leafs whose content on disk is invalid - either garbage
3644          * or the content of some node/leaf from a past generation that got
3645          * cowed or deleted and is no longer valid.
3646          *
3647          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3648          * not be enough - we need to distinguish between log tree extents vs
3649          * non-log tree extents, and the next filemap_fdatawait_range() call
3650          * will catch and clear such errors in the mapping - and that call might
3651          * be from a log sync and not from a transaction commit. Also, checking
3652          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3653          * not done and would not be reliable - the eb might have been released
3654          * from memory and reading it back again means that flag would not be
3655          * set (since it's a runtime flag, not persisted on disk).
3656          *
3657          * Using the flags below in the btree inode also makes us achieve the
3658          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3659          * writeback for all dirty pages and before filemap_fdatawait_range()
3660          * is called, the writeback for all dirty pages had already finished
3661          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3662          * filemap_fdatawait_range() would return success, as it could not know
3663          * that writeback errors happened (the pages were no longer tagged for
3664          * writeback).
3665          */
3666         switch (eb->log_index) {
3667         case -1:
3668                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3669                 break;
3670         case 0:
3671                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3672                 break;
3673         case 1:
3674                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3675                 break;
3676         default:
3677                 BUG(); /* unexpected, logic error */
3678         }
3679 }
3680
3681 static void end_bio_extent_buffer_writepage(struct bio *bio)
3682 {
3683         struct bio_vec *bvec;
3684         struct extent_buffer *eb;
3685         int i, done;
3686
3687         bio_for_each_segment_all(bvec, bio, i) {
3688                 struct page *page = bvec->bv_page;
3689
3690                 eb = (struct extent_buffer *)page->private;
3691                 BUG_ON(!eb);
3692                 done = atomic_dec_and_test(&eb->io_pages);
3693
3694                 if (bio->bi_error ||
3695                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3696                         ClearPageUptodate(page);
3697                         set_btree_ioerr(page);
3698                 }
3699
3700                 end_page_writeback(page);
3701
3702                 if (!done)
3703                         continue;
3704
3705                 end_extent_buffer_writeback(eb);
3706         }
3707
3708         bio_put(bio);
3709 }
3710
3711 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3712                         struct btrfs_fs_info *fs_info,
3713                         struct writeback_control *wbc,
3714                         struct extent_page_data *epd)
3715 {
3716         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3717         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3718         u64 offset = eb->start;
3719         unsigned long i, num_pages;
3720         unsigned long bio_flags = 0;
3721         int write_flags = (epd->sync_io ? WRITE_SYNC : 0) | REQ_META;
3722         int ret = 0;
3723
3724         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3725         num_pages = num_extent_pages(eb->start, eb->len);
3726         atomic_set(&eb->io_pages, num_pages);
3727         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3728                 bio_flags = EXTENT_BIO_TREE_LOG;
3729
3730         /* set btree node beyond nritems with 0 to avoid stale content */
3731         if (btrfs_header_level(eb) > 0) {
3732                 u32 nritems;
3733                 unsigned long end;
3734
3735                 nritems = btrfs_header_nritems(eb);
3736                 end = btrfs_node_key_ptr_offset(nritems);
3737
3738                 memset_extent_buffer(eb, 0, end, eb->len - end);
3739         }
3740
3741         for (i = 0; i < num_pages; i++) {
3742                 struct page *p = eb->pages[i];
3743
3744                 clear_page_dirty_for_io(p);
3745                 set_page_writeback(p);
3746                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3747                                          p, offset >> 9, PAGE_SIZE, 0, bdev,
3748                                          &epd->bio, -1,
3749                                          end_bio_extent_buffer_writepage,
3750                                          0, epd->bio_flags, bio_flags, false);
3751                 epd->bio_flags = bio_flags;
3752                 if (ret) {
3753                         set_btree_ioerr(p);
3754                         end_page_writeback(p);
3755                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3756                                 end_extent_buffer_writeback(eb);
3757                         ret = -EIO;
3758                         break;
3759                 }
3760                 offset += PAGE_SIZE;
3761                 update_nr_written(p, wbc, 1);
3762                 unlock_page(p);
3763         }
3764
3765         if (unlikely(ret)) {
3766                 for (; i < num_pages; i++) {
3767                         struct page *p = eb->pages[i];
3768                         clear_page_dirty_for_io(p);
3769                         unlock_page(p);
3770                 }
3771         }
3772
3773         return ret;
3774 }
3775
3776 int btree_write_cache_pages(struct address_space *mapping,
3777                                    struct writeback_control *wbc)
3778 {
3779         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3780         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3781         struct extent_buffer *eb, *prev_eb = NULL;
3782         struct extent_page_data epd = {
3783                 .bio = NULL,
3784                 .tree = tree,
3785                 .extent_locked = 0,
3786                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3787                 .bio_flags = 0,
3788         };
3789         int ret = 0;
3790         int done = 0;
3791         int nr_to_write_done = 0;
3792         struct pagevec pvec;
3793         int nr_pages;
3794         pgoff_t index;
3795         pgoff_t end;            /* Inclusive */
3796         int scanned = 0;
3797         int tag;
3798
3799         pagevec_init(&pvec, 0);
3800         if (wbc->range_cyclic) {
3801                 index = mapping->writeback_index; /* Start from prev offset */
3802                 end = -1;
3803         } else {
3804                 index = wbc->range_start >> PAGE_SHIFT;
3805                 end = wbc->range_end >> PAGE_SHIFT;
3806                 scanned = 1;
3807         }
3808         if (wbc->sync_mode == WB_SYNC_ALL)
3809                 tag = PAGECACHE_TAG_TOWRITE;
3810         else
3811                 tag = PAGECACHE_TAG_DIRTY;
3812 retry:
3813         if (wbc->sync_mode == WB_SYNC_ALL)
3814                 tag_pages_for_writeback(mapping, index, end);
3815         while (!done && !nr_to_write_done && (index <= end) &&
3816                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3817                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3818                 unsigned i;
3819
3820                 scanned = 1;
3821                 for (i = 0; i < nr_pages; i++) {
3822                         struct page *page = pvec.pages[i];
3823
3824                         if (!PagePrivate(page))
3825                                 continue;
3826
3827                         if (!wbc->range_cyclic && page->index > end) {
3828                                 done = 1;
3829                                 break;
3830                         }
3831
3832                         spin_lock(&mapping->private_lock);
3833                         if (!PagePrivate(page)) {
3834                                 spin_unlock(&mapping->private_lock);
3835                                 continue;
3836                         }
3837
3838                         eb = (struct extent_buffer *)page->private;
3839
3840                         /*
3841                          * Shouldn't happen and normally this would be a BUG_ON
3842                          * but no sense in crashing the users box for something
3843                          * we can survive anyway.
3844                          */
3845                         if (WARN_ON(!eb)) {
3846                                 spin_unlock(&mapping->private_lock);
3847                                 continue;
3848                         }
3849
3850                         if (eb == prev_eb) {
3851                                 spin_unlock(&mapping->private_lock);
3852                                 continue;
3853                         }
3854
3855                         ret = atomic_inc_not_zero(&eb->refs);
3856                         spin_unlock(&mapping->private_lock);
3857                         if (!ret)
3858                                 continue;
3859
3860                         prev_eb = eb;
3861                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3862                         if (!ret) {
3863                                 free_extent_buffer(eb);
3864                                 continue;
3865                         }
3866
3867                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3868                         if (ret) {
3869                                 done = 1;
3870                                 free_extent_buffer(eb);
3871                                 break;
3872                         }
3873                         free_extent_buffer(eb);
3874
3875                         /*
3876                          * the filesystem may choose to bump up nr_to_write.
3877                          * We have to make sure to honor the new nr_to_write
3878                          * at any time
3879                          */
3880                         nr_to_write_done = wbc->nr_to_write <= 0;
3881                 }
3882                 pagevec_release(&pvec);
3883                 cond_resched();
3884         }
3885         if (!scanned && !done) {
3886                 /*
3887                  * We hit the last page and there is more work to be done: wrap
3888                  * back to the start of the file
3889                  */
3890                 scanned = 1;
3891                 index = 0;
3892                 goto retry;
3893         }
3894         flush_write_bio(&epd);
3895         return ret;
3896 }
3897
3898 /**
3899  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3900  * @mapping: address space structure to write
3901  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3902  * @writepage: function called for each page
3903  * @data: data passed to writepage function
3904  *
3905  * If a page is already under I/O, write_cache_pages() skips it, even
3906  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3907  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3908  * and msync() need to guarantee that all the data which was dirty at the time
3909  * the call was made get new I/O started against them.  If wbc->sync_mode is
3910  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3911  * existing IO to complete.
3912  */
3913 static int extent_write_cache_pages(struct extent_io_tree *tree,
3914                              struct address_space *mapping,
3915                              struct writeback_control *wbc,
3916                              writepage_t writepage, void *data,
3917                              void (*flush_fn)(void *))
3918 {
3919         struct inode *inode = mapping->host;
3920         int ret = 0;
3921         int done = 0;
3922         int nr_to_write_done = 0;
3923         struct pagevec pvec;
3924         int nr_pages;
3925         pgoff_t index;
3926         pgoff_t end;            /* Inclusive */
3927         pgoff_t done_index;
3928         int range_whole = 0;
3929         int scanned = 0;
3930         int tag;
3931
3932         /*
3933          * We have to hold onto the inode so that ordered extents can do their
3934          * work when the IO finishes.  The alternative to this is failing to add
3935          * an ordered extent if the igrab() fails there and that is a huge pain
3936          * to deal with, so instead just hold onto the inode throughout the
3937          * writepages operation.  If it fails here we are freeing up the inode
3938          * anyway and we'd rather not waste our time writing out stuff that is
3939          * going to be truncated anyway.
3940          */
3941         if (!igrab(inode))
3942                 return 0;
3943
3944         pagevec_init(&pvec, 0);
3945         if (wbc->range_cyclic) {
3946                 index = mapping->writeback_index; /* Start from prev offset */
3947                 end = -1;
3948         } else {
3949                 index = wbc->range_start >> PAGE_SHIFT;
3950                 end = wbc->range_end >> PAGE_SHIFT;
3951                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3952                         range_whole = 1;
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         done_index = index;
3963         while (!done && !nr_to_write_done && (index <= end) &&
3964                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3965                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3966                 unsigned i;
3967
3968                 scanned = 1;
3969                 for (i = 0; i < nr_pages; i++) {
3970                         struct page *page = pvec.pages[i];
3971
3972                         done_index = page->index;
3973                         /*
3974                          * At this point we hold neither mapping->tree_lock nor
3975                          * lock on the page itself: the page may be truncated or
3976                          * invalidated (changing page->mapping to NULL), or even
3977                          * swizzled back from swapper_space to tmpfs file
3978                          * mapping
3979                          */
3980                         if (!trylock_page(page)) {
3981                                 flush_fn(data);
3982                                 lock_page(page);
3983                         }
3984
3985                         if (unlikely(page->mapping != mapping)) {
3986                                 unlock_page(page);
3987                                 continue;
3988                         }
3989
3990                         if (!wbc->range_cyclic && page->index > end) {
3991                                 done = 1;
3992                                 unlock_page(page);
3993                                 continue;
3994                         }
3995
3996                         if (wbc->sync_mode != WB_SYNC_NONE) {
3997                                 if (PageWriteback(page))
3998                                         flush_fn(data);
3999                                 wait_on_page_writeback(page);
4000                         }
4001
4002                         if (PageWriteback(page) ||
4003                             !clear_page_dirty_for_io(page)) {
4004                                 unlock_page(page);
4005                                 continue;
4006                         }
4007
4008                         ret = (*writepage)(page, wbc, data);
4009
4010                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4011                                 unlock_page(page);
4012                                 ret = 0;
4013                         }
4014                         if (ret < 0) {
4015                                 /*
4016                                  * done_index is set past this page,
4017                                  * so media errors will not choke
4018                                  * background writeout for the entire
4019                                  * file. This has consequences for
4020                                  * range_cyclic semantics (ie. it may
4021                                  * not be suitable for data integrity
4022                                  * writeout).
4023                                  */
4024                                 done_index = page->index + 1;
4025                                 done = 1;
4026                                 break;
4027                         }
4028
4029                         /*
4030                          * the filesystem may choose to bump up nr_to_write.
4031                          * We have to make sure to honor the new nr_to_write
4032                          * at any time
4033                          */
4034                         nr_to_write_done = wbc->nr_to_write <= 0;
4035                 }
4036                 pagevec_release(&pvec);
4037                 cond_resched();
4038         }
4039         if (!scanned && !done) {
4040                 /*
4041                  * We hit the last page and there is more work to be done: wrap
4042                  * back to the start of the file
4043                  */
4044                 scanned = 1;
4045                 index = 0;
4046                 goto retry;
4047         }
4048
4049         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4050                 mapping->writeback_index = done_index;
4051
4052         btrfs_add_delayed_iput(inode);
4053         return ret;
4054 }
4055
4056 static void flush_epd_write_bio(struct extent_page_data *epd)
4057 {
4058         if (epd->bio) {
4059                 int ret;
4060
4061                 bio_set_op_attrs(epd->bio, REQ_OP_WRITE,
4062                                  epd->sync_io ? WRITE_SYNC : 0);
4063
4064                 ret = submit_one_bio(epd->bio, 0, epd->bio_flags);
4065                 BUG_ON(ret < 0); /* -ENOMEM */
4066                 epd->bio = NULL;
4067         }
4068 }
4069
4070 static noinline void flush_write_bio(void *data)
4071 {
4072         struct extent_page_data *epd = data;
4073         flush_epd_write_bio(epd);
4074 }
4075
4076 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4077                           get_extent_t *get_extent,
4078                           struct writeback_control *wbc)
4079 {
4080         int ret;
4081         struct extent_page_data epd = {
4082                 .bio = NULL,
4083                 .tree = tree,
4084                 .get_extent = get_extent,
4085                 .extent_locked = 0,
4086                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4087                 .bio_flags = 0,
4088         };
4089
4090         ret = __extent_writepage(page, wbc, &epd);
4091
4092         flush_epd_write_bio(&epd);
4093         return ret;
4094 }
4095
4096 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4097                               u64 start, u64 end, get_extent_t *get_extent,
4098                               int mode)
4099 {
4100         int ret = 0;
4101         struct address_space *mapping = inode->i_mapping;
4102         struct page *page;
4103         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4104                 PAGE_SHIFT;
4105
4106         struct extent_page_data epd = {
4107                 .bio = NULL,
4108                 .tree = tree,
4109                 .get_extent = get_extent,
4110                 .extent_locked = 1,
4111                 .sync_io = mode == WB_SYNC_ALL,
4112                 .bio_flags = 0,
4113         };
4114         struct writeback_control wbc_writepages = {
4115                 .sync_mode      = mode,
4116                 .nr_to_write    = nr_pages * 2,
4117                 .range_start    = start,
4118                 .range_end      = end + 1,
4119         };
4120
4121         while (start <= end) {
4122                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4123                 if (clear_page_dirty_for_io(page))
4124                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4125                 else {
4126                         if (tree->ops && tree->ops->writepage_end_io_hook)
4127                                 tree->ops->writepage_end_io_hook(page, start,
4128                                                  start + PAGE_SIZE - 1,
4129                                                  NULL, 1);
4130                         unlock_page(page);
4131                 }
4132                 put_page(page);
4133                 start += PAGE_SIZE;
4134         }
4135
4136         flush_epd_write_bio(&epd);
4137         return ret;
4138 }
4139
4140 int extent_writepages(struct extent_io_tree *tree,
4141                       struct address_space *mapping,
4142                       get_extent_t *get_extent,
4143                       struct writeback_control *wbc)
4144 {
4145         int ret = 0;
4146         struct extent_page_data epd = {
4147                 .bio = NULL,
4148                 .tree = tree,
4149                 .get_extent = get_extent,
4150                 .extent_locked = 0,
4151                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4152                 .bio_flags = 0,
4153         };
4154
4155         ret = extent_write_cache_pages(tree, mapping, wbc,
4156                                        __extent_writepage, &epd,
4157                                        flush_write_bio);
4158         flush_epd_write_bio(&epd);
4159         return ret;
4160 }
4161
4162 int extent_readpages(struct extent_io_tree *tree,
4163                      struct address_space *mapping,
4164                      struct list_head *pages, unsigned nr_pages,
4165                      get_extent_t get_extent)
4166 {
4167         struct bio *bio = NULL;
4168         unsigned page_idx;
4169         unsigned long bio_flags = 0;
4170         struct page *pagepool[16];
4171         struct page *page;
4172         struct extent_map *em_cached = NULL;
4173         int nr = 0;
4174         u64 prev_em_start = (u64)-1;
4175
4176         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4177                 page = list_entry(pages->prev, struct page, lru);
4178
4179                 prefetchw(&page->flags);
4180                 list_del(&page->lru);
4181                 if (add_to_page_cache_lru(page, mapping,
4182                                         page->index,
4183                                         readahead_gfp_mask(mapping))) {
4184                         put_page(page);
4185                         continue;
4186                 }
4187
4188                 pagepool[nr++] = page;
4189                 if (nr < ARRAY_SIZE(pagepool))
4190                         continue;
4191                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4192                                    &bio, 0, &bio_flags, &prev_em_start);
4193                 nr = 0;
4194         }
4195         if (nr)
4196                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4197                                    &bio, 0, &bio_flags, &prev_em_start);
4198
4199         if (em_cached)
4200                 free_extent_map(em_cached);
4201
4202         BUG_ON(!list_empty(pages));
4203         if (bio)
4204                 return submit_one_bio(bio, 0, bio_flags);
4205         return 0;
4206 }
4207
4208 /*
4209  * basic invalidatepage code, this waits on any locked or writeback
4210  * ranges corresponding to the page, and then deletes any extent state
4211  * records from the tree
4212  */
4213 int extent_invalidatepage(struct extent_io_tree *tree,
4214                           struct page *page, unsigned long offset)
4215 {
4216         struct extent_state *cached_state = NULL;
4217         u64 start = page_offset(page);
4218         u64 end = start + PAGE_SIZE - 1;
4219         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4220
4221         start += ALIGN(offset, blocksize);
4222         if (start > end)
4223                 return 0;
4224
4225         lock_extent_bits(tree, start, end, &cached_state);
4226         wait_on_page_writeback(page);
4227         clear_extent_bit(tree, start, end,
4228                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4229                          EXTENT_DO_ACCOUNTING,
4230                          1, 1, &cached_state, GFP_NOFS);
4231         return 0;
4232 }
4233
4234 /*
4235  * a helper for releasepage, this tests for areas of the page that
4236  * are locked or under IO and drops the related state bits if it is safe
4237  * to drop the page.
4238  */
4239 static int try_release_extent_state(struct extent_map_tree *map,
4240                                     struct extent_io_tree *tree,
4241                                     struct page *page, gfp_t mask)
4242 {
4243         u64 start = page_offset(page);
4244         u64 end = start + PAGE_SIZE - 1;
4245         int ret = 1;
4246
4247         if (test_range_bit(tree, start, end,
4248                            EXTENT_IOBITS, 0, NULL))
4249                 ret = 0;
4250         else {
4251                 if ((mask & GFP_NOFS) == GFP_NOFS)
4252                         mask = GFP_NOFS;
4253                 /*
4254                  * at this point we can safely clear everything except the
4255                  * locked bit and the nodatasum bit
4256                  */
4257                 ret = clear_extent_bit(tree, start, end,
4258                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4259                                  0, 0, NULL, mask);
4260
4261                 /* if clear_extent_bit failed for enomem reasons,
4262                  * we can't allow the release to continue.
4263                  */
4264                 if (ret < 0)
4265                         ret = 0;
4266                 else
4267                         ret = 1;
4268         }
4269         return ret;
4270 }
4271
4272 /*
4273  * a helper for releasepage.  As long as there are no locked extents
4274  * in the range corresponding to the page, both state records and extent
4275  * map records are removed
4276  */
4277 int try_release_extent_mapping(struct extent_map_tree *map,
4278                                struct extent_io_tree *tree, struct page *page,
4279                                gfp_t mask)
4280 {
4281         struct extent_map *em;
4282         u64 start = page_offset(page);
4283         u64 end = start + PAGE_SIZE - 1;
4284
4285         if (gfpflags_allow_blocking(mask) &&
4286             page->mapping->host->i_size > SZ_16M) {
4287                 u64 len;
4288                 while (start <= end) {
4289                         len = end - start + 1;
4290                         write_lock(&map->lock);
4291                         em = lookup_extent_mapping(map, start, len);
4292                         if (!em) {
4293                                 write_unlock(&map->lock);
4294                                 break;
4295                         }
4296                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4297                             em->start != start) {
4298                                 write_unlock(&map->lock);
4299                                 free_extent_map(em);
4300                                 break;
4301                         }
4302                         if (!test_range_bit(tree, em->start,
4303                                             extent_map_end(em) - 1,
4304                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4305                                             0, NULL)) {
4306                                 remove_extent_mapping(map, em);
4307                                 /* once for the rb tree */
4308                                 free_extent_map(em);
4309                         }
4310                         start = extent_map_end(em);
4311                         write_unlock(&map->lock);
4312
4313                         /* once for us */
4314                         free_extent_map(em);
4315                 }
4316         }
4317         return try_release_extent_state(map, tree, page, mask);
4318 }
4319
4320 /*
4321  * helper function for fiemap, which doesn't want to see any holes.
4322  * This maps until we find something past 'last'
4323  */
4324 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4325                                                 u64 offset,
4326                                                 u64 last,
4327                                                 get_extent_t *get_extent)
4328 {
4329         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4330         struct extent_map *em;
4331         u64 len;
4332
4333         if (offset >= last)
4334                 return NULL;
4335
4336         while (1) {
4337                 len = last - offset;
4338                 if (len == 0)
4339                         break;
4340                 len = ALIGN(len, sectorsize);
4341                 em = get_extent(inode, NULL, 0, offset, len, 0);
4342                 if (IS_ERR_OR_NULL(em))
4343                         return em;
4344
4345                 /* if this isn't a hole return it */
4346                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4347                     em->block_start != EXTENT_MAP_HOLE) {
4348                         return em;
4349                 }
4350
4351                 /* this is a hole, advance to the next extent */
4352                 offset = extent_map_end(em);
4353                 free_extent_map(em);
4354                 if (offset >= last)
4355                         break;
4356         }
4357         return NULL;
4358 }
4359
4360 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4361                 __u64 start, __u64 len, get_extent_t *get_extent)
4362 {
4363         int ret = 0;
4364         u64 off = start;
4365         u64 max = start + len;
4366         u32 flags = 0;
4367         u32 found_type;
4368         u64 last;
4369         u64 last_for_get_extent = 0;
4370         u64 disko = 0;
4371         u64 isize = i_size_read(inode);
4372         struct btrfs_key found_key;
4373         struct extent_map *em = NULL;
4374         struct extent_state *cached_state = NULL;
4375         struct btrfs_path *path;
4376         struct btrfs_root *root = BTRFS_I(inode)->root;
4377         int end = 0;
4378         u64 em_start = 0;
4379         u64 em_len = 0;
4380         u64 em_end = 0;
4381
4382         if (len == 0)
4383                 return -EINVAL;
4384
4385         path = btrfs_alloc_path();
4386         if (!path)
4387                 return -ENOMEM;
4388         path->leave_spinning = 1;
4389
4390         start = round_down(start, BTRFS_I(inode)->root->sectorsize);
4391         len = round_up(max, BTRFS_I(inode)->root->sectorsize) - start;
4392
4393         /*
4394          * lookup the last file extent.  We're not using i_size here
4395          * because there might be preallocation past i_size
4396          */
4397         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode), -1,
4398                                        0);
4399         if (ret < 0) {
4400                 btrfs_free_path(path);
4401                 return ret;
4402         } else {
4403                 WARN_ON(!ret);
4404                 if (ret == 1)
4405                         ret = 0;
4406         }
4407
4408         path->slots[0]--;
4409         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4410         found_type = found_key.type;
4411
4412         /* No extents, but there might be delalloc bits */
4413         if (found_key.objectid != btrfs_ino(inode) ||
4414             found_type != BTRFS_EXTENT_DATA_KEY) {
4415                 /* have to trust i_size as the end */
4416                 last = (u64)-1;
4417                 last_for_get_extent = isize;
4418         } else {
4419                 /*
4420                  * remember the start of the last extent.  There are a
4421                  * bunch of different factors that go into the length of the
4422                  * extent, so its much less complex to remember where it started
4423                  */
4424                 last = found_key.offset;
4425                 last_for_get_extent = last + 1;
4426         }
4427         btrfs_release_path(path);
4428
4429         /*
4430          * we might have some extents allocated but more delalloc past those
4431          * extents.  so, we trust isize unless the start of the last extent is
4432          * beyond isize
4433          */
4434         if (last < isize) {
4435                 last = (u64)-1;
4436                 last_for_get_extent = isize;
4437         }
4438
4439         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4440                          &cached_state);
4441
4442         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4443                                    get_extent);
4444         if (!em)
4445                 goto out;
4446         if (IS_ERR(em)) {
4447                 ret = PTR_ERR(em);
4448                 goto out;
4449         }
4450
4451         while (!end) {
4452                 u64 offset_in_extent = 0;
4453
4454                 /* break if the extent we found is outside the range */
4455                 if (em->start >= max || extent_map_end(em) < off)
4456                         break;
4457
4458                 /*
4459                  * get_extent may return an extent that starts before our
4460                  * requested range.  We have to make sure the ranges
4461                  * we return to fiemap always move forward and don't
4462                  * overlap, so adjust the offsets here
4463                  */
4464                 em_start = max(em->start, off);
4465
4466                 /*
4467                  * record the offset from the start of the extent
4468                  * for adjusting the disk offset below.  Only do this if the
4469                  * extent isn't compressed since our in ram offset may be past
4470                  * what we have actually allocated on disk.
4471                  */
4472                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4473                         offset_in_extent = em_start - em->start;
4474                 em_end = extent_map_end(em);
4475                 em_len = em_end - em_start;
4476                 disko = 0;
4477                 flags = 0;
4478
4479                 /*
4480                  * bump off for our next call to get_extent
4481                  */
4482                 off = extent_map_end(em);
4483                 if (off >= max)
4484                         end = 1;
4485
4486                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4487                         end = 1;
4488                         flags |= FIEMAP_EXTENT_LAST;
4489                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4490                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4491                                   FIEMAP_EXTENT_NOT_ALIGNED);
4492                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4493                         flags |= (FIEMAP_EXTENT_DELALLOC |
4494                                   FIEMAP_EXTENT_UNKNOWN);
4495                 } else if (fieinfo->fi_extents_max) {
4496                         struct btrfs_trans_handle *trans;
4497
4498                         u64 bytenr = em->block_start -
4499                                 (em->start - em->orig_start);
4500
4501                         disko = em->block_start + offset_in_extent;
4502
4503                         /*
4504                          * We need a trans handle to get delayed refs
4505                          */
4506                         trans = btrfs_join_transaction(root);
4507                         /*
4508                          * It's OK if we can't start a trans we can still check
4509                          * from commit_root
4510                          */
4511                         if (IS_ERR(trans))
4512                                 trans = NULL;
4513
4514                         /*
4515                          * As btrfs supports shared space, this information
4516                          * can be exported to userspace tools via
4517                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4518                          * then we're just getting a count and we can skip the
4519                          * lookup stuff.
4520                          */
4521                         ret = btrfs_check_shared(trans, root->fs_info,
4522                                                  root->objectid,
4523                                                  btrfs_ino(inode), bytenr);
4524                         if (trans)
4525                                 btrfs_end_transaction(trans, root);
4526                         if (ret < 0)
4527                                 goto out_free;
4528                         if (ret)
4529                                 flags |= FIEMAP_EXTENT_SHARED;
4530                         ret = 0;
4531                 }
4532                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4533                         flags |= FIEMAP_EXTENT_ENCODED;
4534                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4535                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4536
4537                 free_extent_map(em);
4538                 em = NULL;
4539                 if ((em_start >= last) || em_len == (u64)-1 ||
4540                    (last == (u64)-1 && isize <= em_end)) {
4541                         flags |= FIEMAP_EXTENT_LAST;
4542                         end = 1;
4543                 }
4544
4545                 /* now scan forward to see if this is really the last extent. */
4546                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4547                                            get_extent);
4548                 if (IS_ERR(em)) {
4549                         ret = PTR_ERR(em);
4550                         goto out;
4551                 }
4552                 if (!em) {
4553                         flags |= FIEMAP_EXTENT_LAST;
4554                         end = 1;
4555                 }
4556                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4557                                               em_len, flags);
4558                 if (ret) {
4559                         if (ret == 1)
4560                                 ret = 0;
4561                         goto out_free;
4562                 }
4563         }
4564 out_free:
4565         free_extent_map(em);
4566 out:
4567         btrfs_free_path(path);
4568         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4569                              &cached_state, GFP_NOFS);
4570         return ret;
4571 }
4572
4573 static void __free_extent_buffer(struct extent_buffer *eb)
4574 {
4575         btrfs_leak_debug_del(&eb->leak_list);
4576         kmem_cache_free(extent_buffer_cache, eb);
4577 }
4578
4579 int extent_buffer_under_io(struct extent_buffer *eb)
4580 {
4581         return (atomic_read(&eb->io_pages) ||
4582                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4583                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4584 }
4585
4586 /*
4587  * Helper for releasing extent buffer page.
4588  */
4589 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4590 {
4591         unsigned long index;
4592         struct page *page;
4593         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4594
4595         BUG_ON(extent_buffer_under_io(eb));
4596
4597         index = num_extent_pages(eb->start, eb->len);
4598         if (index == 0)
4599                 return;
4600
4601         do {
4602                 index--;
4603                 page = eb->pages[index];
4604                 if (!page)
4605                         continue;
4606                 if (mapped)
4607                         spin_lock(&page->mapping->private_lock);
4608                 /*
4609                  * We do this since we'll remove the pages after we've
4610                  * removed the eb from the radix tree, so we could race
4611                  * and have this page now attached to the new eb.  So
4612                  * only clear page_private if it's still connected to
4613                  * this eb.
4614                  */
4615                 if (PagePrivate(page) &&
4616                     page->private == (unsigned long)eb) {
4617                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4618                         BUG_ON(PageDirty(page));
4619                         BUG_ON(PageWriteback(page));
4620                         /*
4621                          * We need to make sure we haven't be attached
4622                          * to a new eb.
4623                          */
4624                         ClearPagePrivate(page);
4625                         set_page_private(page, 0);
4626                         /* One for the page private */
4627                         put_page(page);
4628                 }
4629
4630                 if (mapped)
4631                         spin_unlock(&page->mapping->private_lock);
4632
4633                 /* One for when we allocated the page */
4634                 put_page(page);
4635         } while (index != 0);
4636 }
4637
4638 /*
4639  * Helper for releasing the extent buffer.
4640  */
4641 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4642 {
4643         btrfs_release_extent_buffer_page(eb);
4644         __free_extent_buffer(eb);
4645 }
4646
4647 static struct extent_buffer *
4648 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4649                       unsigned long len)
4650 {
4651         struct extent_buffer *eb = NULL;
4652
4653         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4654         eb->start = start;
4655         eb->len = len;
4656         eb->fs_info = fs_info;
4657         eb->bflags = 0;
4658         rwlock_init(&eb->lock);
4659         atomic_set(&eb->write_locks, 0);
4660         atomic_set(&eb->read_locks, 0);
4661         atomic_set(&eb->blocking_readers, 0);
4662         atomic_set(&eb->blocking_writers, 0);
4663         atomic_set(&eb->spinning_readers, 0);
4664         atomic_set(&eb->spinning_writers, 0);
4665         eb->lock_nested = 0;
4666         init_waitqueue_head(&eb->write_lock_wq);
4667         init_waitqueue_head(&eb->read_lock_wq);
4668
4669         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4670
4671         spin_lock_init(&eb->refs_lock);
4672         atomic_set(&eb->refs, 1);
4673         atomic_set(&eb->io_pages, 0);
4674
4675         /*
4676          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4677          */
4678         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4679                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4680         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4681
4682         return eb;
4683 }
4684
4685 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4686 {
4687         unsigned long i;
4688         struct page *p;
4689         struct extent_buffer *new;
4690         unsigned long num_pages = num_extent_pages(src->start, src->len);
4691
4692         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4693         if (new == NULL)
4694                 return NULL;
4695
4696         for (i = 0; i < num_pages; i++) {
4697                 p = alloc_page(GFP_NOFS);
4698                 if (!p) {
4699                         btrfs_release_extent_buffer(new);
4700                         return NULL;
4701                 }
4702                 attach_extent_buffer_page(new, p);
4703                 WARN_ON(PageDirty(p));
4704                 SetPageUptodate(p);
4705                 new->pages[i] = p;
4706         }
4707
4708         copy_extent_buffer(new, src, 0, 0, src->len);
4709         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4710         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4711
4712         return new;
4713 }
4714
4715 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4716                                                   u64 start, unsigned long len)
4717 {
4718         struct extent_buffer *eb;
4719         unsigned long num_pages;
4720         unsigned long i;
4721
4722         num_pages = num_extent_pages(start, len);
4723
4724         eb = __alloc_extent_buffer(fs_info, start, len);
4725         if (!eb)
4726                 return NULL;
4727
4728         for (i = 0; i < num_pages; i++) {
4729                 eb->pages[i] = alloc_page(GFP_NOFS);
4730                 if (!eb->pages[i])
4731                         goto err;
4732         }
4733         set_extent_buffer_uptodate(eb);
4734         btrfs_set_header_nritems(eb, 0);
4735         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4736
4737         return eb;
4738 err:
4739         for (; i > 0; i--)
4740                 __free_page(eb->pages[i - 1]);
4741         __free_extent_buffer(eb);
4742         return NULL;
4743 }
4744
4745 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4746                                                 u64 start, u32 nodesize)
4747 {
4748         unsigned long len;
4749
4750         if (!fs_info) {
4751                 /*
4752                  * Called only from tests that don't always have a fs_info
4753                  * available
4754                  */
4755                 len = nodesize;
4756         } else {
4757                 len = fs_info->tree_root->nodesize;
4758         }
4759
4760         return __alloc_dummy_extent_buffer(fs_info, start, len);
4761 }
4762
4763 static void check_buffer_tree_ref(struct extent_buffer *eb)
4764 {
4765         int refs;
4766         /* the ref bit is tricky.  We have to make sure it is set
4767          * if we have the buffer dirty.   Otherwise the
4768          * code to free a buffer can end up dropping a dirty
4769          * page
4770          *
4771          * Once the ref bit is set, it won't go away while the
4772          * buffer is dirty or in writeback, and it also won't
4773          * go away while we have the reference count on the
4774          * eb bumped.
4775          *
4776          * We can't just set the ref bit without bumping the
4777          * ref on the eb because free_extent_buffer might
4778          * see the ref bit and try to clear it.  If this happens
4779          * free_extent_buffer might end up dropping our original
4780          * ref by mistake and freeing the page before we are able
4781          * to add one more ref.
4782          *
4783          * So bump the ref count first, then set the bit.  If someone
4784          * beat us to it, drop the ref we added.
4785          */
4786         refs = atomic_read(&eb->refs);
4787         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4788                 return;
4789
4790         spin_lock(&eb->refs_lock);
4791         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4792                 atomic_inc(&eb->refs);
4793         spin_unlock(&eb->refs_lock);
4794 }
4795
4796 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4797                 struct page *accessed)
4798 {
4799         unsigned long num_pages, i;
4800
4801         check_buffer_tree_ref(eb);
4802
4803         num_pages = num_extent_pages(eb->start, eb->len);
4804         for (i = 0; i < num_pages; i++) {
4805                 struct page *p = eb->pages[i];
4806
4807                 if (p != accessed)
4808                         mark_page_accessed(p);
4809         }
4810 }
4811
4812 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4813                                          u64 start)
4814 {
4815         struct extent_buffer *eb;
4816
4817         rcu_read_lock();
4818         eb = radix_tree_lookup(&fs_info->buffer_radix,
4819                                start >> PAGE_SHIFT);
4820         if (eb && atomic_inc_not_zero(&eb->refs)) {
4821                 rcu_read_unlock();
4822                 /*
4823                  * Lock our eb's refs_lock to avoid races with
4824                  * free_extent_buffer. When we get our eb it might be flagged
4825                  * with EXTENT_BUFFER_STALE and another task running
4826                  * free_extent_buffer might have seen that flag set,
4827                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4828                  * writeback flags not set) and it's still in the tree (flag
4829                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4830                  * of decrementing the extent buffer's reference count twice.
4831                  * So here we could race and increment the eb's reference count,
4832                  * clear its stale flag, mark it as dirty and drop our reference
4833                  * before the other task finishes executing free_extent_buffer,
4834                  * which would later result in an attempt to free an extent
4835                  * buffer that is dirty.
4836                  */
4837                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4838                         spin_lock(&eb->refs_lock);
4839                         spin_unlock(&eb->refs_lock);
4840                 }
4841                 mark_extent_buffer_accessed(eb, NULL);
4842                 return eb;
4843         }
4844         rcu_read_unlock();
4845
4846         return NULL;
4847 }
4848
4849 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4850 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4851                                         u64 start, u32 nodesize)
4852 {
4853         struct extent_buffer *eb, *exists = NULL;
4854         int ret;
4855
4856         eb = find_extent_buffer(fs_info, start);
4857         if (eb)
4858                 return eb;
4859         eb = alloc_dummy_extent_buffer(fs_info, start, nodesize);
4860         if (!eb)
4861                 return NULL;
4862         eb->fs_info = fs_info;
4863 again:
4864         ret = radix_tree_preload(GFP_NOFS);
4865         if (ret)
4866                 goto free_eb;
4867         spin_lock(&fs_info->buffer_lock);
4868         ret = radix_tree_insert(&fs_info->buffer_radix,
4869                                 start >> PAGE_SHIFT, eb);
4870         spin_unlock(&fs_info->buffer_lock);
4871         radix_tree_preload_end();
4872         if (ret == -EEXIST) {
4873                 exists = find_extent_buffer(fs_info, start);
4874                 if (exists)
4875                         goto free_eb;
4876                 else
4877                         goto again;
4878         }
4879         check_buffer_tree_ref(eb);
4880         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4881
4882         /*
4883          * We will free dummy extent buffer's if they come into
4884          * free_extent_buffer with a ref count of 2, but if we are using this we
4885          * want the buffers to stay in memory until we're done with them, so
4886          * bump the ref count again.
4887          */
4888         atomic_inc(&eb->refs);
4889         return eb;
4890 free_eb:
4891         btrfs_release_extent_buffer(eb);
4892         return exists;
4893 }
4894 #endif
4895
4896 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4897                                           u64 start)
4898 {
4899         unsigned long len = fs_info->tree_root->nodesize;
4900         unsigned long num_pages = num_extent_pages(start, len);
4901         unsigned long i;
4902         unsigned long index = start >> PAGE_SHIFT;
4903         struct extent_buffer *eb;
4904         struct extent_buffer *exists = NULL;
4905         struct page *p;
4906         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4907         int uptodate = 1;
4908         int ret;
4909
4910         if (!IS_ALIGNED(start, fs_info->tree_root->sectorsize)) {
4911                 btrfs_err(fs_info, "bad tree block start %llu", start);
4912                 return ERR_PTR(-EINVAL);
4913         }
4914
4915         eb = find_extent_buffer(fs_info, start);
4916         if (eb)
4917                 return eb;
4918
4919         eb = __alloc_extent_buffer(fs_info, start, len);
4920         if (!eb)
4921                 return ERR_PTR(-ENOMEM);
4922
4923         for (i = 0; i < num_pages; i++, index++) {
4924                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4925                 if (!p) {
4926                         exists = ERR_PTR(-ENOMEM);
4927                         goto free_eb;
4928                 }
4929
4930                 spin_lock(&mapping->private_lock);
4931                 if (PagePrivate(p)) {
4932                         /*
4933                          * We could have already allocated an eb for this page
4934                          * and attached one so lets see if we can get a ref on
4935                          * the existing eb, and if we can we know it's good and
4936                          * we can just return that one, else we know we can just
4937                          * overwrite page->private.
4938                          */
4939                         exists = (struct extent_buffer *)p->private;
4940                         if (atomic_inc_not_zero(&exists->refs)) {
4941                                 spin_unlock(&mapping->private_lock);
4942                                 unlock_page(p);
4943                                 put_page(p);
4944                                 mark_extent_buffer_accessed(exists, p);
4945                                 goto free_eb;
4946                         }
4947                         exists = NULL;
4948
4949                         /*
4950                          * Do this so attach doesn't complain and we need to
4951                          * drop the ref the old guy had.
4952                          */
4953                         ClearPagePrivate(p);
4954                         WARN_ON(PageDirty(p));
4955                         put_page(p);
4956                 }
4957                 attach_extent_buffer_page(eb, p);
4958                 spin_unlock(&mapping->private_lock);
4959                 WARN_ON(PageDirty(p));
4960                 eb->pages[i] = p;
4961                 if (!PageUptodate(p))
4962                         uptodate = 0;
4963
4964                 /*
4965                  * see below about how we avoid a nasty race with release page
4966                  * and why we unlock later
4967                  */
4968         }
4969         if (uptodate)
4970                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4971 again:
4972         ret = radix_tree_preload(GFP_NOFS);
4973         if (ret) {
4974                 exists = ERR_PTR(ret);
4975                 goto free_eb;
4976         }
4977
4978         spin_lock(&fs_info->buffer_lock);
4979         ret = radix_tree_insert(&fs_info->buffer_radix,
4980                                 start >> PAGE_SHIFT, eb);
4981         spin_unlock(&fs_info->buffer_lock);
4982         radix_tree_preload_end();
4983         if (ret == -EEXIST) {
4984                 exists = find_extent_buffer(fs_info, start);
4985                 if (exists)
4986                         goto free_eb;
4987                 else
4988                         goto again;
4989         }
4990         /* add one reference for the tree */
4991         check_buffer_tree_ref(eb);
4992         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4993
4994         /*
4995          * there is a race where release page may have
4996          * tried to find this extent buffer in the radix
4997          * but failed.  It will tell the VM it is safe to
4998          * reclaim the, and it will clear the page private bit.
4999          * We must make sure to set the page private bit properly
5000          * after the extent buffer is in the radix tree so
5001          * it doesn't get lost
5002          */
5003         SetPageChecked(eb->pages[0]);
5004         for (i = 1; i < num_pages; i++) {
5005                 p = eb->pages[i];
5006                 ClearPageChecked(p);
5007                 unlock_page(p);
5008         }
5009         unlock_page(eb->pages[0]);
5010         return eb;
5011
5012 free_eb:
5013         WARN_ON(!atomic_dec_and_test(&eb->refs));
5014         for (i = 0; i < num_pages; i++) {
5015                 if (eb->pages[i])
5016                         unlock_page(eb->pages[i]);
5017         }
5018
5019         btrfs_release_extent_buffer(eb);
5020         return exists;
5021 }
5022
5023 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5024 {
5025         struct extent_buffer *eb =
5026                         container_of(head, struct extent_buffer, rcu_head);
5027
5028         __free_extent_buffer(eb);
5029 }
5030
5031 /* Expects to have eb->eb_lock already held */
5032 static int release_extent_buffer(struct extent_buffer *eb)
5033 {
5034         WARN_ON(atomic_read(&eb->refs) == 0);
5035         if (atomic_dec_and_test(&eb->refs)) {
5036                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5037                         struct btrfs_fs_info *fs_info = eb->fs_info;
5038
5039                         spin_unlock(&eb->refs_lock);
5040
5041                         spin_lock(&fs_info->buffer_lock);
5042                         radix_tree_delete(&fs_info->buffer_radix,
5043                                           eb->start >> PAGE_SHIFT);
5044                         spin_unlock(&fs_info->buffer_lock);
5045                 } else {
5046                         spin_unlock(&eb->refs_lock);
5047                 }
5048
5049                 /* Should be safe to release our pages at this point */
5050                 btrfs_release_extent_buffer_page(eb);
5051 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5052                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5053                         __free_extent_buffer(eb);
5054                         return 1;
5055                 }
5056 #endif
5057                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5058                 return 1;
5059         }
5060         spin_unlock(&eb->refs_lock);
5061
5062         return 0;
5063 }
5064
5065 void free_extent_buffer(struct extent_buffer *eb)
5066 {
5067         int refs;
5068         int old;
5069         if (!eb)
5070                 return;
5071
5072         while (1) {
5073                 refs = atomic_read(&eb->refs);
5074                 if (refs <= 3)
5075                         break;
5076                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5077                 if (old == refs)
5078                         return;
5079         }
5080
5081         spin_lock(&eb->refs_lock);
5082         if (atomic_read(&eb->refs) == 2 &&
5083             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5084                 atomic_dec(&eb->refs);
5085
5086         if (atomic_read(&eb->refs) == 2 &&
5087             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5088             !extent_buffer_under_io(eb) &&
5089             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5090                 atomic_dec(&eb->refs);
5091
5092         /*
5093          * I know this is terrible, but it's temporary until we stop tracking
5094          * the uptodate bits and such for the extent buffers.
5095          */
5096         release_extent_buffer(eb);
5097 }
5098
5099 void free_extent_buffer_stale(struct extent_buffer *eb)
5100 {
5101         if (!eb)
5102                 return;
5103
5104         spin_lock(&eb->refs_lock);
5105         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5106
5107         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5108             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5109                 atomic_dec(&eb->refs);
5110         release_extent_buffer(eb);
5111 }
5112
5113 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5114 {
5115         unsigned long i;
5116         unsigned long num_pages;
5117         struct page *page;
5118
5119         num_pages = num_extent_pages(eb->start, eb->len);
5120
5121         for (i = 0; i < num_pages; i++) {
5122                 page = eb->pages[i];
5123                 if (!PageDirty(page))
5124                         continue;
5125
5126                 lock_page(page);
5127                 WARN_ON(!PagePrivate(page));
5128
5129                 clear_page_dirty_for_io(page);
5130                 spin_lock_irq(&page->mapping->tree_lock);
5131                 if (!PageDirty(page)) {
5132                         radix_tree_tag_clear(&page->mapping->page_tree,
5133                                                 page_index(page),
5134                                                 PAGECACHE_TAG_DIRTY);
5135                 }
5136                 spin_unlock_irq(&page->mapping->tree_lock);
5137                 ClearPageError(page);
5138                 unlock_page(page);
5139         }
5140         WARN_ON(atomic_read(&eb->refs) == 0);
5141 }
5142
5143 int set_extent_buffer_dirty(struct extent_buffer *eb)
5144 {
5145         unsigned long i;
5146         unsigned long num_pages;
5147         int was_dirty = 0;
5148
5149         check_buffer_tree_ref(eb);
5150
5151         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5152
5153         num_pages = num_extent_pages(eb->start, eb->len);
5154         WARN_ON(atomic_read(&eb->refs) == 0);
5155         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5156
5157         for (i = 0; i < num_pages; i++)
5158                 set_page_dirty(eb->pages[i]);
5159         return was_dirty;
5160 }
5161
5162 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5163 {
5164         unsigned long i;
5165         struct page *page;
5166         unsigned long num_pages;
5167
5168         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5169         num_pages = num_extent_pages(eb->start, eb->len);
5170         for (i = 0; i < num_pages; i++) {
5171                 page = eb->pages[i];
5172                 if (page)
5173                         ClearPageUptodate(page);
5174         }
5175 }
5176
5177 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5178 {
5179         unsigned long i;
5180         struct page *page;
5181         unsigned long num_pages;
5182
5183         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5184         num_pages = num_extent_pages(eb->start, eb->len);
5185         for (i = 0; i < num_pages; i++) {
5186                 page = eb->pages[i];
5187                 SetPageUptodate(page);
5188         }
5189 }
5190
5191 int extent_buffer_uptodate(struct extent_buffer *eb)
5192 {
5193         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5194 }
5195
5196 int read_extent_buffer_pages(struct extent_io_tree *tree,
5197                              struct extent_buffer *eb, int wait,
5198                              get_extent_t *get_extent, int mirror_num)
5199 {
5200         unsigned long i;
5201         struct page *page;
5202         int err;
5203         int ret = 0;
5204         int locked_pages = 0;
5205         int all_uptodate = 1;
5206         unsigned long num_pages;
5207         unsigned long num_reads = 0;
5208         struct bio *bio = NULL;
5209         unsigned long bio_flags = 0;
5210
5211         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5212                 return 0;
5213
5214         num_pages = num_extent_pages(eb->start, eb->len);
5215         for (i = 0; i < num_pages; i++) {
5216                 page = eb->pages[i];
5217                 if (wait == WAIT_NONE) {
5218                         if (!trylock_page(page))
5219                                 goto unlock_exit;
5220                 } else {
5221                         lock_page(page);
5222                 }
5223                 locked_pages++;
5224         }
5225         /*
5226          * We need to firstly lock all pages to make sure that
5227          * the uptodate bit of our pages won't be affected by
5228          * clear_extent_buffer_uptodate().
5229          */
5230         for (i = 0; i < num_pages; i++) {
5231                 page = eb->pages[i];
5232                 if (!PageUptodate(page)) {
5233                         num_reads++;
5234                         all_uptodate = 0;
5235                 }
5236         }
5237
5238         if (all_uptodate) {
5239                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5240                 goto unlock_exit;
5241         }
5242
5243         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5244         eb->read_mirror = 0;
5245         atomic_set(&eb->io_pages, num_reads);
5246         for (i = 0; i < num_pages; i++) {
5247                 page = eb->pages[i];
5248
5249                 if (!PageUptodate(page)) {
5250                         if (ret) {
5251                                 atomic_dec(&eb->io_pages);
5252                                 unlock_page(page);
5253                                 continue;
5254                         }
5255
5256                         ClearPageError(page);
5257                         err = __extent_read_full_page(tree, page,
5258                                                       get_extent, &bio,
5259                                                       mirror_num, &bio_flags,
5260                                                       REQ_META);
5261                         if (err) {
5262                                 ret = err;
5263                                 /*
5264                                  * We use &bio in above __extent_read_full_page,
5265                                  * so we ensure that if it returns error, the
5266                                  * current page fails to add itself to bio and
5267                                  * it's been unlocked.
5268                                  *
5269                                  * We must dec io_pages by ourselves.
5270                                  */
5271                                 atomic_dec(&eb->io_pages);
5272                         }
5273                 } else {
5274                         unlock_page(page);
5275                 }
5276         }
5277
5278         if (bio) {
5279                 err = submit_one_bio(bio, mirror_num, bio_flags);
5280                 if (err)
5281                         return err;
5282         }
5283
5284         if (ret || wait != WAIT_COMPLETE)
5285                 return ret;
5286
5287         for (i = 0; i < num_pages; i++) {
5288                 page = eb->pages[i];
5289                 wait_on_page_locked(page);
5290                 if (!PageUptodate(page))
5291                         ret = -EIO;
5292         }
5293
5294         return ret;
5295
5296 unlock_exit:
5297         while (locked_pages > 0) {
5298                 locked_pages--;
5299                 page = eb->pages[locked_pages];
5300                 unlock_page(page);
5301         }
5302         return ret;
5303 }
5304
5305 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5306                         unsigned long start,
5307                         unsigned long len)
5308 {
5309         size_t cur;
5310         size_t offset;
5311         struct page *page;
5312         char *kaddr;
5313         char *dst = (char *)dstv;
5314         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5315         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5316
5317         WARN_ON(start > eb->len);
5318         WARN_ON(start + len > eb->start + eb->len);
5319
5320         offset = (start_offset + start) & (PAGE_SIZE - 1);
5321
5322         while (len > 0) {
5323                 page = eb->pages[i];
5324
5325                 cur = min(len, (PAGE_SIZE - offset));
5326                 kaddr = page_address(page);
5327                 memcpy(dst, kaddr + offset, cur);
5328
5329                 dst += cur;
5330                 len -= cur;
5331                 offset = 0;
5332                 i++;
5333         }
5334 }
5335
5336 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5337                         unsigned long start,
5338                         unsigned long len)
5339 {
5340         size_t cur;
5341         size_t offset;
5342         struct page *page;
5343         char *kaddr;
5344         char __user *dst = (char __user *)dstv;
5345         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5346         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5347         int ret = 0;
5348
5349         WARN_ON(start > eb->len);
5350         WARN_ON(start + len > eb->start + eb->len);
5351
5352         offset = (start_offset + start) & (PAGE_SIZE - 1);
5353
5354         while (len > 0) {
5355                 page = eb->pages[i];
5356
5357                 cur = min(len, (PAGE_SIZE - offset));
5358                 kaddr = page_address(page);
5359                 if (copy_to_user(dst, kaddr + offset, cur)) {
5360                         ret = -EFAULT;
5361                         break;
5362                 }
5363
5364                 dst += cur;
5365                 len -= cur;
5366                 offset = 0;
5367                 i++;
5368         }
5369
5370         return ret;
5371 }
5372
5373 /*
5374  * return 0 if the item is found within a page.
5375  * return 1 if the item spans two pages.
5376  * return -EINVAL otherwise.
5377  */
5378 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5379                                unsigned long min_len, char **map,
5380                                unsigned long *map_start,
5381                                unsigned long *map_len)
5382 {
5383         size_t offset = start & (PAGE_SIZE - 1);
5384         char *kaddr;
5385         struct page *p;
5386         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5387         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5388         unsigned long end_i = (start_offset + start + min_len - 1) >>
5389                 PAGE_SHIFT;
5390
5391         if (i != end_i)
5392                 return 1;
5393
5394         if (i == 0) {
5395                 offset = start_offset;
5396                 *map_start = 0;
5397         } else {
5398                 offset = 0;
5399                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5400         }
5401
5402         if (start + min_len > eb->len) {
5403                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5404                        eb->start, eb->len, start, min_len);
5405                 return -EINVAL;
5406         }
5407
5408         p = eb->pages[i];
5409         kaddr = page_address(p);
5410         *map = kaddr + offset;
5411         *map_len = PAGE_SIZE - offset;
5412         return 0;
5413 }
5414
5415 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5416                           unsigned long start,
5417                           unsigned long len)
5418 {
5419         size_t cur;
5420         size_t offset;
5421         struct page *page;
5422         char *kaddr;
5423         char *ptr = (char *)ptrv;
5424         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5425         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5426         int ret = 0;
5427
5428         WARN_ON(start > eb->len);
5429         WARN_ON(start + len > eb->start + eb->len);
5430
5431         offset = (start_offset + start) & (PAGE_SIZE - 1);
5432
5433         while (len > 0) {
5434                 page = eb->pages[i];
5435
5436                 cur = min(len, (PAGE_SIZE - offset));
5437
5438                 kaddr = page_address(page);
5439                 ret = memcmp(ptr, kaddr + offset, cur);
5440                 if (ret)
5441                         break;
5442
5443                 ptr += cur;
5444                 len -= cur;
5445                 offset = 0;
5446                 i++;
5447         }
5448         return ret;
5449 }
5450
5451 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5452                          unsigned long start, unsigned long len)
5453 {
5454         size_t cur;
5455         size_t offset;
5456         struct page *page;
5457         char *kaddr;
5458         char *src = (char *)srcv;
5459         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5460         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5461
5462         WARN_ON(start > eb->len);
5463         WARN_ON(start + len > eb->start + eb->len);
5464
5465         offset = (start_offset + start) & (PAGE_SIZE - 1);
5466
5467         while (len > 0) {
5468                 page = eb->pages[i];
5469                 WARN_ON(!PageUptodate(page));
5470
5471                 cur = min(len, PAGE_SIZE - offset);
5472                 kaddr = page_address(page);
5473                 memcpy(kaddr + offset, src, cur);
5474
5475                 src += cur;
5476                 len -= cur;
5477                 offset = 0;
5478                 i++;
5479         }
5480 }
5481
5482 void memset_extent_buffer(struct extent_buffer *eb, char c,
5483                           unsigned long start, unsigned long len)
5484 {
5485         size_t cur;
5486         size_t offset;
5487         struct page *page;
5488         char *kaddr;
5489         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5490         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5491
5492         WARN_ON(start > eb->len);
5493         WARN_ON(start + len > eb->start + eb->len);
5494
5495         offset = (start_offset + start) & (PAGE_SIZE - 1);
5496
5497         while (len > 0) {
5498                 page = eb->pages[i];
5499                 WARN_ON(!PageUptodate(page));
5500
5501                 cur = min(len, PAGE_SIZE - offset);
5502                 kaddr = page_address(page);
5503                 memset(kaddr + offset, c, cur);
5504
5505                 len -= cur;
5506                 offset = 0;
5507                 i++;
5508         }
5509 }
5510
5511 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5512                         unsigned long dst_offset, unsigned long src_offset,
5513                         unsigned long len)
5514 {
5515         u64 dst_len = dst->len;
5516         size_t cur;
5517         size_t offset;
5518         struct page *page;
5519         char *kaddr;
5520         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5521         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5522
5523         WARN_ON(src->len != dst_len);
5524
5525         offset = (start_offset + dst_offset) &
5526                 (PAGE_SIZE - 1);
5527
5528         while (len > 0) {
5529                 page = dst->pages[i];
5530                 WARN_ON(!PageUptodate(page));
5531
5532                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5533
5534                 kaddr = page_address(page);
5535                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5536
5537                 src_offset += cur;
5538                 len -= cur;
5539                 offset = 0;
5540                 i++;
5541         }
5542 }
5543
5544 /*
5545  * The extent buffer bitmap operations are done with byte granularity because
5546  * bitmap items are not guaranteed to be aligned to a word and therefore a
5547  * single word in a bitmap may straddle two pages in the extent buffer.
5548  */
5549 #define BIT_BYTE(nr) ((nr) / BITS_PER_BYTE)
5550 #define BYTE_MASK ((1 << BITS_PER_BYTE) - 1)
5551 #define BITMAP_FIRST_BYTE_MASK(start) \
5552         ((BYTE_MASK << ((start) & (BITS_PER_BYTE - 1))) & BYTE_MASK)
5553 #define BITMAP_LAST_BYTE_MASK(nbits) \
5554         (BYTE_MASK >> (-(nbits) & (BITS_PER_BYTE - 1)))
5555
5556 /*
5557  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5558  * given bit number
5559  * @eb: the extent buffer
5560  * @start: offset of the bitmap item in the extent buffer
5561  * @nr: bit number
5562  * @page_index: return index of the page in the extent buffer that contains the
5563  * given bit number
5564  * @page_offset: return offset into the page given by page_index
5565  *
5566  * This helper hides the ugliness of finding the byte in an extent buffer which
5567  * contains a given bit.
5568  */
5569 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5570                                     unsigned long start, unsigned long nr,
5571                                     unsigned long *page_index,
5572                                     size_t *page_offset)
5573 {
5574         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5575         size_t byte_offset = BIT_BYTE(nr);
5576         size_t offset;
5577
5578         /*
5579          * The byte we want is the offset of the extent buffer + the offset of
5580          * the bitmap item in the extent buffer + the offset of the byte in the
5581          * bitmap item.
5582          */
5583         offset = start_offset + start + byte_offset;
5584
5585         *page_index = offset >> PAGE_SHIFT;
5586         *page_offset = offset & (PAGE_SIZE - 1);
5587 }
5588
5589 /**
5590  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5591  * @eb: the extent buffer
5592  * @start: offset of the bitmap item in the extent buffer
5593  * @nr: bit number to test
5594  */
5595 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5596                            unsigned long nr)
5597 {
5598         char *kaddr;
5599         struct page *page;
5600         unsigned long i;
5601         size_t offset;
5602
5603         eb_bitmap_offset(eb, start, nr, &i, &offset);
5604         page = eb->pages[i];
5605         WARN_ON(!PageUptodate(page));
5606         kaddr = page_address(page);
5607         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5608 }
5609
5610 /**
5611  * extent_buffer_bitmap_set - set an area of a bitmap
5612  * @eb: the extent buffer
5613  * @start: offset of the bitmap item in the extent buffer
5614  * @pos: bit number of the first bit
5615  * @len: number of bits to set
5616  */
5617 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5618                               unsigned long pos, unsigned long len)
5619 {
5620         char *kaddr;
5621         struct page *page;
5622         unsigned long i;
5623         size_t offset;
5624         const unsigned int size = pos + len;
5625         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5626         unsigned int mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5627
5628         eb_bitmap_offset(eb, start, pos, &i, &offset);
5629         page = eb->pages[i];
5630         WARN_ON(!PageUptodate(page));
5631         kaddr = page_address(page);
5632
5633         while (len >= bits_to_set) {
5634                 kaddr[offset] |= mask_to_set;
5635                 len -= bits_to_set;
5636                 bits_to_set = BITS_PER_BYTE;
5637                 mask_to_set = ~0U;
5638                 if (++offset >= PAGE_SIZE && len > 0) {
5639                         offset = 0;
5640                         page = eb->pages[++i];
5641                         WARN_ON(!PageUptodate(page));
5642                         kaddr = page_address(page);
5643                 }
5644         }
5645         if (len) {
5646                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5647                 kaddr[offset] |= mask_to_set;
5648         }
5649 }
5650
5651
5652 /**
5653  * extent_buffer_bitmap_clear - clear an area of a bitmap
5654  * @eb: the extent buffer
5655  * @start: offset of the bitmap item in the extent buffer
5656  * @pos: bit number of the first bit
5657  * @len: number of bits to clear
5658  */
5659 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5660                                 unsigned long pos, unsigned long len)
5661 {
5662         char *kaddr;
5663         struct page *page;
5664         unsigned long i;
5665         size_t offset;
5666         const unsigned int size = pos + len;
5667         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5668         unsigned int mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5669
5670         eb_bitmap_offset(eb, start, pos, &i, &offset);
5671         page = eb->pages[i];
5672         WARN_ON(!PageUptodate(page));
5673         kaddr = page_address(page);
5674
5675         while (len >= bits_to_clear) {
5676                 kaddr[offset] &= ~mask_to_clear;
5677                 len -= bits_to_clear;
5678                 bits_to_clear = BITS_PER_BYTE;
5679                 mask_to_clear = ~0U;
5680                 if (++offset >= PAGE_SIZE && len > 0) {
5681                         offset = 0;
5682                         page = eb->pages[++i];
5683                         WARN_ON(!PageUptodate(page));
5684                         kaddr = page_address(page);
5685                 }
5686         }
5687         if (len) {
5688                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5689                 kaddr[offset] &= ~mask_to_clear;
5690         }
5691 }
5692
5693 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5694 {
5695         unsigned long distance = (src > dst) ? src - dst : dst - src;
5696         return distance < len;
5697 }
5698
5699 static void copy_pages(struct page *dst_page, struct page *src_page,
5700                        unsigned long dst_off, unsigned long src_off,
5701                        unsigned long len)
5702 {
5703         char *dst_kaddr = page_address(dst_page);
5704         char *src_kaddr;
5705         int must_memmove = 0;
5706
5707         if (dst_page != src_page) {
5708                 src_kaddr = page_address(src_page);
5709         } else {
5710                 src_kaddr = dst_kaddr;
5711                 if (areas_overlap(src_off, dst_off, len))
5712                         must_memmove = 1;
5713         }
5714
5715         if (must_memmove)
5716                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5717         else
5718                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5719 }
5720
5721 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5722                            unsigned long src_offset, unsigned long len)
5723 {
5724         size_t cur;
5725         size_t dst_off_in_page;
5726         size_t src_off_in_page;
5727         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5728         unsigned long dst_i;
5729         unsigned long src_i;
5730
5731         if (src_offset + len > dst->len) {
5732                 btrfs_err(dst->fs_info,
5733                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5734                          src_offset, len, dst->len);
5735                 BUG_ON(1);
5736         }
5737         if (dst_offset + len > dst->len) {
5738                 btrfs_err(dst->fs_info,
5739                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5740                          dst_offset, len, dst->len);
5741                 BUG_ON(1);
5742         }
5743
5744         while (len > 0) {
5745                 dst_off_in_page = (start_offset + dst_offset) &
5746                         (PAGE_SIZE - 1);
5747                 src_off_in_page = (start_offset + src_offset) &
5748                         (PAGE_SIZE - 1);
5749
5750                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5751                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5752
5753                 cur = min(len, (unsigned long)(PAGE_SIZE -
5754                                                src_off_in_page));
5755                 cur = min_t(unsigned long, cur,
5756                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5757
5758                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5759                            dst_off_in_page, src_off_in_page, cur);
5760
5761                 src_offset += cur;
5762                 dst_offset += cur;
5763                 len -= cur;
5764         }
5765 }
5766
5767 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5768                            unsigned long src_offset, unsigned long len)
5769 {
5770         size_t cur;
5771         size_t dst_off_in_page;
5772         size_t src_off_in_page;
5773         unsigned long dst_end = dst_offset + len - 1;
5774         unsigned long src_end = src_offset + len - 1;
5775         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5776         unsigned long dst_i;
5777         unsigned long src_i;
5778
5779         if (src_offset + len > dst->len) {
5780                 btrfs_err(dst->fs_info,
5781                           "memmove bogus src_offset %lu move len %lu len %lu",
5782                           src_offset, len, dst->len);
5783                 BUG_ON(1);
5784         }
5785         if (dst_offset + len > dst->len) {
5786                 btrfs_err(dst->fs_info,
5787                           "memmove bogus dst_offset %lu move len %lu len %lu",
5788                           dst_offset, len, dst->len);
5789                 BUG_ON(1);
5790         }
5791         if (dst_offset < src_offset) {
5792                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5793                 return;
5794         }
5795         while (len > 0) {
5796                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5797                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5798
5799                 dst_off_in_page = (start_offset + dst_end) &
5800                         (PAGE_SIZE - 1);
5801                 src_off_in_page = (start_offset + src_end) &
5802                         (PAGE_SIZE - 1);
5803
5804                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5805                 cur = min(cur, dst_off_in_page + 1);
5806                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5807                            dst_off_in_page - cur + 1,
5808                            src_off_in_page - cur + 1, cur);
5809
5810                 dst_end -= cur;
5811                 src_end -= cur;
5812                 len -= cur;
5813         }
5814 }
5815
5816 int try_release_extent_buffer(struct page *page)
5817 {
5818         struct extent_buffer *eb;
5819
5820         /*
5821          * We need to make sure nobody is attaching this page to an eb right
5822          * now.
5823          */
5824         spin_lock(&page->mapping->private_lock);
5825         if (!PagePrivate(page)) {
5826                 spin_unlock(&page->mapping->private_lock);
5827                 return 1;
5828         }
5829
5830         eb = (struct extent_buffer *)page->private;
5831         BUG_ON(!eb);
5832
5833         /*
5834          * This is a little awful but should be ok, we need to make sure that
5835          * the eb doesn't disappear out from under us while we're looking at
5836          * this page.
5837          */
5838         spin_lock(&eb->refs_lock);
5839         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5840                 spin_unlock(&eb->refs_lock);
5841                 spin_unlock(&page->mapping->private_lock);
5842                 return 0;
5843         }
5844         spin_unlock(&page->mapping->private_lock);
5845
5846         /*
5847          * If tree ref isn't set then we know the ref on this eb is a real ref,
5848          * so just return, this page will likely be freed soon anyway.
5849          */
5850         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5851                 spin_unlock(&eb->refs_lock);
5852                 return 0;
5853         }
5854
5855         return release_extent_buffer(eb);
5856 }