fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/spinlock.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/swap.h>
  10 #include <linux/writeback.h>
  11 #include <linux/pagevec.h>
  12 #include <linux/prefetch.h>
  13 #include <linux/cleancache.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "ctree.h"
  17 #include "btrfs_inode.h"
  18 #include "volumes.h"
  19 #include "check-integrity.h"
  20 #include "locking.h"
  21 #include "rcu-string.h"
  22 #include "backref.h"
  23 #include "transaction.h"
  24
  25 static struct kmem_cache *extent_state_cache;
  26 static struct kmem_cache *extent_buffer_cache;
  27 static struct bio_set *btrfs_bioset;
  28
  29 static inline bool extent_state_in_tree(const struct extent_state *state)
  30 {
  31         return !RB_EMPTY_NODE(&state->rb_node);
  32 }
  33
  34 #ifdef CONFIG_BTRFS_DEBUG
  35 static LIST_HEAD(buffers);
  36 static LIST_HEAD(states);
  37
  38 static DEFINE_SPINLOCK(leak_lock);
  39
  40 static inline
  41 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  42 {
  43         unsigned long flags;
  44
  45         spin_lock_irqsave(&leak_lock, flags);
  46         list_add(new, head);
  47         spin_unlock_irqrestore(&leak_lock, flags);
  48 }
  49
  50 static inline
  51 void btrfs_leak_debug_del(struct list_head *entry)
  52 {
  53         unsigned long flags;
  54
  55         spin_lock_irqsave(&leak_lock, flags);
  56         list_del(entry);
  57         spin_unlock_irqrestore(&leak_lock, flags);
  58 }
  59
  60 static inline
  61 void btrfs_leak_debug_check(void)
  62 {
  63         struct extent_state *state;
  64         struct extent_buffer *eb;
  65
  66         while (!list_empty(&states)) {
  67                 state = list_entry(states.next, struct extent_state, leak_list);
  68                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  69                        state->start, state->end, state->state,
  70                        extent_state_in_tree(state),
  71                        atomic_read(&state->refs));
  72                 list_del(&state->leak_list);
  73                 kmem_cache_free(extent_state_cache, state);
  74         }
  75
  76         while (!list_empty(&buffers)) {
  77                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  78                 pr_err("BTRFS: buffer leak start %llu len %lu refs %d\n",
  79                        eb->start, eb->len, atomic_read(&eb->refs));
  80                 list_del(&eb->leak_list);
  81                 kmem_cache_free(extent_buffer_cache, eb);
  82         }
  83 }
  84
  85 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  86         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  87 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  88                 struct extent_io_tree *tree, u64 start, u64 end)
  89 {
  90         struct inode *inode;
  91         u64 isize;
  92
  93         if (!tree->mapping)
  94                 return;
  95
  96         inode = tree->mapping->host;
  97         isize = i_size_read(inode);
  98         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
  99                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
 100                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
 101                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
 102         }
 103 }
 104 #else
 105 #define btrfs_leak_debug_add(new, head) do {} while (0)
 106 #define btrfs_leak_debug_del(entry)     do {} while (0)
 107 #define btrfs_leak_debug_check()        do {} while (0)
 108 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 109 #endif
 110
 111 #define BUFFER_LRU_MAX 64
 112
 113 struct tree_entry {
 114         u64 start;
 115         u64 end;
 116         struct rb_node rb_node;
 117 };
 118
 119 struct extent_page_data {
 120         struct bio *bio;
 121         struct extent_io_tree *tree;
 122         get_extent_t *get_extent;
 123         unsigned long bio_flags;
 124
 125         /* tells writepage not to lock the state bits for this range
 126          * it still does the unlocking
 127          */
 128         unsigned int extent_locked:1;
 129
 130         /* tells the submit_bio code to use REQ_SYNC */
 131         unsigned int sync_io:1;
 132 };
 133
 134 static void add_extent_changeset(struct extent_state *state, unsigned bits,
 135                                  struct extent_changeset *changeset,
 136                                  int set)
 137 {
 138         int ret;
 139
 140         if (!changeset)
 141                 return;
 142         if (set && (state->state & bits) == bits)
 143                 return;
 144         if (!set && (state->state & bits) == 0)
 145                 return;
 146         changeset->bytes_changed += state->end - state->start + 1;
 147         ret = ulist_add(&changeset->range_changed, state->start, state->end,
 148                         GFP_ATOMIC);
 149         /* ENOMEM */
 150         BUG_ON(ret < 0);
 151 }
 152
 153 static noinline void flush_write_bio(void *data);
 154 static inline struct btrfs_fs_info *
 155 tree_fs_info(struct extent_io_tree *tree)
 156 {
 157         if (!tree->mapping)
 158                 return NULL;
 159         return btrfs_sb(tree->mapping->host->i_sb);
 160 }
 161
 162 int __init extent_io_init(void)
 163 {
 164         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 165                         sizeof(struct extent_state), 0,
 166                         SLAB_MEM_SPREAD, NULL);
 167         if (!extent_state_cache)
 168                 return -ENOMEM;
 169
 170         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 171                         sizeof(struct extent_buffer), 0,
 172                         SLAB_MEM_SPREAD, NULL);
 173         if (!extent_buffer_cache)
 174                 goto free_state_cache;
 175
 176         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 177                                      offsetof(struct btrfs_io_bio, bio));
 178         if (!btrfs_bioset)
 179                 goto free_buffer_cache;
 180
 181         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 182                 goto free_bioset;
 183
 184         return 0;
 185
 186 free_bioset:
 187         bioset_free(btrfs_bioset);
 188         btrfs_bioset = NULL;
 189
 190 free_buffer_cache:
 191         kmem_cache_destroy(extent_buffer_cache);
 192         extent_buffer_cache = NULL;
 193
 194 free_state_cache:
 195         kmem_cache_destroy(extent_state_cache);
 196         extent_state_cache = NULL;
 197         return -ENOMEM;
 198 }
 199
 200 void extent_io_exit(void)
 201 {
 202         btrfs_leak_debug_check();
 203
 204         /*
 205          * Make sure all delayed rcu free are flushed before we
 206          * destroy caches.
 207          */
 208         rcu_barrier();
 209         kmem_cache_destroy(extent_state_cache);
 210         kmem_cache_destroy(extent_buffer_cache);
 211         if (btrfs_bioset)
 212                 bioset_free(btrfs_bioset);
 213 }
 214
 215 void extent_io_tree_init(struct extent_io_tree *tree,
 216                          struct address_space *mapping)
 217 {
 218         tree->state = RB_ROOT;
 219         tree->ops = NULL;
 220         tree->dirty_bytes = 0;
 221         spin_lock_init(&tree->lock);
 222         tree->mapping = mapping;
 223 }
 224
 225 static struct extent_state *alloc_extent_state(gfp_t mask)
 226 {
 227         struct extent_state *state;
 228
 229         /*
 230          * The given mask might be not appropriate for the slab allocator,
 231          * drop the unsupported bits
 232          */
 233         mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
 234         state = kmem_cache_alloc(extent_state_cache, mask);
 235         if (!state)
 236                 return state;
 237         state->state = 0;
 238         state->failrec = NULL;
 239         RB_CLEAR_NODE(&state->rb_node);
 240         btrfs_leak_debug_add(&state->leak_list, &states);
 241         atomic_set(&state->refs, 1);
 242         init_waitqueue_head(&state->wq);
 243         trace_alloc_extent_state(state, mask, _RET_IP_);
 244         return state;
 245 }
 246
 247 void free_extent_state(struct extent_state *state)
 248 {
 249         if (!state)
 250                 return;
 251         if (atomic_dec_and_test(&state->refs)) {
 252                 WARN_ON(extent_state_in_tree(state));
 253                 btrfs_leak_debug_del(&state->leak_list);
 254                 trace_free_extent_state(state, _RET_IP_);
 255                 kmem_cache_free(extent_state_cache, state);
 256         }
 257 }
 258
 259 static struct rb_node *tree_insert(struct rb_root *root,
 260                                    struct rb_node *search_start,
 261                                    u64 offset,
 262                                    struct rb_node *node,
 263                                    struct rb_node ***p_in,
 264                                    struct rb_node **parent_in)
 265 {
 266         struct rb_node **p;
 267         struct rb_node *parent = NULL;
 268         struct tree_entry *entry;
 269
 270         if (p_in && parent_in) {
 271                 p = *p_in;
 272                 parent = *parent_in;
 273                 goto do_insert;
 274         }
 275
 276         p = search_start ? &search_start : &root->rb_node;
 277         while (*p) {
 278                 parent = *p;
 279                 entry = rb_entry(parent, struct tree_entry, rb_node);
 280
 281                 if (offset < entry->start)
 282                         p = &(*p)->rb_left;
 283                 else if (offset > entry->end)
 284                         p = &(*p)->rb_right;
 285                 else
 286                         return parent;
 287         }
 288
 289 do_insert:
 290         rb_link_node(node, parent, p);
 291         rb_insert_color(node, root);
 292         return NULL;
 293 }
 294
 295 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 296                                       struct rb_node **prev_ret,
 297                                       struct rb_node **next_ret,
 298                                       struct rb_node ***p_ret,
 299                                       struct rb_node **parent_ret)
 300 {
 301         struct rb_root *root = &tree->state;
 302         struct rb_node **n = &root->rb_node;
 303         struct rb_node *prev = NULL;
 304         struct rb_node *orig_prev = NULL;
 305         struct tree_entry *entry;
 306         struct tree_entry *prev_entry = NULL;
 307
 308         while (*n) {
 309                 prev = *n;
 310                 entry = rb_entry(prev, struct tree_entry, rb_node);
 311                 prev_entry = entry;
 312
 313                 if (offset < entry->start)
 314                         n = &(*n)->rb_left;
 315                 else if (offset > entry->end)
 316                         n = &(*n)->rb_right;
 317                 else
 318                         return *n;
 319         }
 320
 321         if (p_ret)
 322                 *p_ret = n;
 323         if (parent_ret)
 324                 *parent_ret = prev;
 325
 326         if (prev_ret) {
 327                 orig_prev = prev;
 328                 while (prev && offset > prev_entry->end) {
 329                         prev = rb_next(prev);
 330                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 331                 }
 332                 *prev_ret = prev;
 333                 prev = orig_prev;
 334         }
 335
 336         if (next_ret) {
 337                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 338                 while (prev && offset < prev_entry->start) {
 339                         prev = rb_prev(prev);
 340                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 341                 }
 342                 *next_ret = prev;
 343         }
 344         return NULL;
 345 }
 346
 347 static inline struct rb_node *
 348 tree_search_for_insert(struct extent_io_tree *tree,
 349                        u64 offset,
 350                        struct rb_node ***p_ret,
 351                        struct rb_node **parent_ret)
 352 {
 353         struct rb_node *prev = NULL;
 354         struct rb_node *ret;
 355
 356         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 357         if (!ret)
 358                 return prev;
 359         return ret;
 360 }
 361
 362 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 363                                           u64 offset)
 364 {
 365         return tree_search_for_insert(tree, offset, NULL, NULL);
 366 }
 367
 368 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 369                      struct extent_state *other)
 370 {
 371         if (tree->ops && tree->ops->merge_extent_hook)
 372                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 373                                              other);
 374 }
 375
 376 /*
 377  * utility function to look for merge candidates inside a given range.
 378  * Any extents with matching state are merged together into a single
 379  * extent in the tree.  Extents with EXTENT_IO in their state field
 380  * are not merged because the end_io handlers need to be able to do
 381  * operations on them without sleeping (or doing allocations/splits).
 382  *
 383  * This should be called with the tree lock held.
 384  */
 385 static void merge_state(struct extent_io_tree *tree,
 386                         struct extent_state *state)
 387 {
 388         struct extent_state *other;
 389         struct rb_node *other_node;
 390
 391         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 392                 return;
 393
 394         other_node = rb_prev(&state->rb_node);
 395         if (other_node) {
 396                 other = rb_entry(other_node, struct extent_state, rb_node);
 397                 if (other->end == state->start - 1 &&
 398                     other->state == state->state) {
 399                         merge_cb(tree, state, other);
 400                         state->start = other->start;
 401                         rb_erase(&other->rb_node, &tree->state);
 402                         RB_CLEAR_NODE(&other->rb_node);
 403                         free_extent_state(other);
 404                 }
 405         }
 406         other_node = rb_next(&state->rb_node);
 407         if (other_node) {
 408                 other = rb_entry(other_node, struct extent_state, rb_node);
 409                 if (other->start == state->end + 1 &&
 410                     other->state == state->state) {
 411                         merge_cb(tree, state, other);
 412                         state->end = other->end;
 413                         rb_erase(&other->rb_node, &tree->state);
 414                         RB_CLEAR_NODE(&other->rb_node);
 415                         free_extent_state(other);
 416                 }
 417         }
 418 }
 419
 420 static void set_state_cb(struct extent_io_tree *tree,
 421                          struct extent_state *state, unsigned *bits)
 422 {
 423         if (tree->ops && tree->ops->set_bit_hook)
 424                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 425 }
 426
 427 static void clear_state_cb(struct extent_io_tree *tree,
 428                            struct extent_state *state, unsigned *bits)
 429 {
 430         if (tree->ops && tree->ops->clear_bit_hook)
 431                 tree->ops->clear_bit_hook(BTRFS_I(tree->mapping->host),
 432                                 state, bits);
 433 }
 434
 435 static void set_state_bits(struct extent_io_tree *tree,
 436                            struct extent_state *state, unsigned *bits,
 437                            struct extent_changeset *changeset);
 438
 439 /*
 440  * insert an extent_state struct into the tree.  'bits' are set on the
 441  * struct before it is inserted.
 442  *
 443  * This may return -EEXIST if the extent is already there, in which case the
 444  * state struct is freed.
 445  *
 446  * The tree lock is not taken internally.  This is a utility function and
 447  * probably isn't what you want to call (see set/clear_extent_bit).
 448  */
 449 static int insert_state(struct extent_io_tree *tree,
 450                         struct extent_state *state, u64 start, u64 end,
 451                         struct rb_node ***p,
 452                         struct rb_node **parent,
 453                         unsigned *bits, struct extent_changeset *changeset)
 454 {
 455         struct rb_node *node;
 456
 457         if (end < start)
 458                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 459                        end, start);
 460         state->start = start;
 461         state->end = end;
 462
 463         set_state_bits(tree, state, bits, changeset);
 464
 465         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 466         if (node) {
 467                 struct extent_state *found;
 468                 found = rb_entry(node, struct extent_state, rb_node);
 469                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
 470                        found->start, found->end, start, end);
 471                 return -EEXIST;
 472         }
 473         merge_state(tree, state);
 474         return 0;
 475 }
 476
 477 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 478                      u64 split)
 479 {
 480         if (tree->ops && tree->ops->split_extent_hook)
 481                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 482 }
 483
 484 /*
 485  * split a given extent state struct in two, inserting the preallocated
 486  * struct 'prealloc' as the newly created second half.  'split' indicates an
 487  * offset inside 'orig' where it should be split.
 488  *
 489  * Before calling,
 490  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 491  * are two extent state structs in the tree:
 492  * prealloc: [orig->start, split - 1]
 493  * orig: [ split, orig->end ]
 494  *
 495  * The tree locks are not taken by this function. They need to be held
 496  * by the caller.
 497  */
 498 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 499                        struct extent_state *prealloc, u64 split)
 500 {
 501         struct rb_node *node;
 502
 503         split_cb(tree, orig, split);
 504
 505         prealloc->start = orig->start;
 506         prealloc->end = split - 1;
 507         prealloc->state = orig->state;
 508         orig->start = split;
 509
 510         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 511                            &prealloc->rb_node, NULL, NULL);
 512         if (node) {
 513                 free_extent_state(prealloc);
 514                 return -EEXIST;
 515         }
 516         return 0;
 517 }
 518
 519 static struct extent_state *next_state(struct extent_state *state)
 520 {
 521         struct rb_node *next = rb_next(&state->rb_node);
 522         if (next)
 523                 return rb_entry(next, struct extent_state, rb_node);
 524         else
 525                 return NULL;
 526 }
 527
 528 /*
 529  * utility function to clear some bits in an extent state struct.
 530  * it will optionally wake up any one waiting on this state (wake == 1).
 531  *
 532  * If no bits are set on the state struct after clearing things, the
 533  * struct is freed and removed from the tree
 534  */
 535 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 536                                             struct extent_state *state,
 537                                             unsigned *bits, int wake,
 538                                             struct extent_changeset *changeset)
 539 {
 540         struct extent_state *next;
 541         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 542
 543         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 544                 u64 range = state->end - state->start + 1;
 545                 WARN_ON(range > tree->dirty_bytes);
 546                 tree->dirty_bytes -= range;
 547         }
 548         clear_state_cb(tree, state, bits);
 549         add_extent_changeset(state, bits_to_clear, changeset, 0);
 550         state->state &= ~bits_to_clear;
 551         if (wake)
 552                 wake_up(&state->wq);
 553         if (state->state == 0) {
 554                 next = next_state(state);
 555                 if (extent_state_in_tree(state)) {
 556                         rb_erase(&state->rb_node, &tree->state);
 557                         RB_CLEAR_NODE(&state->rb_node);
 558                         free_extent_state(state);
 559                 } else {
 560                         WARN_ON(1);
 561                 }
 562         } else {
 563                 merge_state(tree, state);
 564                 next = next_state(state);
 565         }
 566         return next;
 567 }
 568
 569 static struct extent_state *
 570 alloc_extent_state_atomic(struct extent_state *prealloc)
 571 {
 572         if (!prealloc)
 573                 prealloc = alloc_extent_state(GFP_ATOMIC);
 574
 575         return prealloc;
 576 }
 577
 578 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 579 {
 580         btrfs_panic(tree_fs_info(tree), err,
 581                     "Locking error: Extent tree was modified by another thread while locked.");
 582 }
 583
 584 /*
 585  * clear some bits on a range in the tree.  This may require splitting
 586  * or inserting elements in the tree, so the gfp mask is used to
 587  * indicate which allocations or sleeping are allowed.
 588  *
 589  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 590  * the given range from the tree regardless of state (ie for truncate).
 591  *
 592  * the range [start, end] is inclusive.
 593  *
 594  * This takes the tree lock, and returns 0 on success and < 0 on error.
 595  */
 596 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 597                               unsigned bits, int wake, int delete,
 598                               struct extent_state **cached_state,
 599                               gfp_t mask, struct extent_changeset *changeset)
 600 {
 601         struct extent_state *state;
 602         struct extent_state *cached;
 603         struct extent_state *prealloc = NULL;
 604         struct rb_node *node;
 605         u64 last_end;
 606         int err;
 607         int clear = 0;
 608
 609         btrfs_debug_check_extent_io_range(tree, start, end);
 610
 611         if (bits & EXTENT_DELALLOC)
 612                 bits |= EXTENT_NORESERVE;
 613
 614         if (delete)
 615                 bits |= ~EXTENT_CTLBITS;
 616         bits |= EXTENT_FIRST_DELALLOC;
 617
 618         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 619                 clear = 1;
 620 again:
 621         if (!prealloc && gfpflags_allow_blocking(mask)) {
 622                 /*
 623                  * Don't care for allocation failure here because we might end
 624                  * up not needing the pre-allocated extent state at all, which
 625                  * is the case if we only have in the tree extent states that
 626                  * cover our input range and don't cover too any other range.
 627                  * If we end up needing a new extent state we allocate it later.
 628                  */
 629                 prealloc = alloc_extent_state(mask);
 630         }
 631
 632         spin_lock(&tree->lock);
 633         if (cached_state) {
 634                 cached = *cached_state;
 635
 636                 if (clear) {
 637                         *cached_state = NULL;
 638                         cached_state = NULL;
 639                 }
 640
 641                 if (cached && extent_state_in_tree(cached) &&
 642                     cached->start <= start && cached->end > start) {
 643                         if (clear)
 644                                 atomic_dec(&cached->refs);
 645                         state = cached;
 646                         goto hit_next;
 647                 }
 648                 if (clear)
 649                         free_extent_state(cached);
 650         }
 651         /*
 652          * this search will find the extents that end after
 653          * our range starts
 654          */
 655         node = tree_search(tree, start);
 656         if (!node)
 657                 goto out;
 658         state = rb_entry(node, struct extent_state, rb_node);
 659 hit_next:
 660         if (state->start > end)
 661                 goto out;
 662         WARN_ON(state->end < start);
 663         last_end = state->end;
 664
 665         /* the state doesn't have the wanted bits, go ahead */
 666         if (!(state->state & bits)) {
 667                 state = next_state(state);
 668                 goto next;
 669         }
 670
 671         /*
 672          *     | ---- desired range ---- |
 673          *  | state | or
 674          *  | ------------- state -------------- |
 675          *
 676          * We need to split the extent we found, and may flip
 677          * bits on second half.
 678          *
 679          * If the extent we found extends past our range, we
 680          * just split and search again.  It'll get split again
 681          * the next time though.
 682          *
 683          * If the extent we found is inside our range, we clear
 684          * the desired bit on it.
 685          */
 686
 687         if (state->start < start) {
 688                 prealloc = alloc_extent_state_atomic(prealloc);
 689                 BUG_ON(!prealloc);
 690                 err = split_state(tree, state, prealloc, start);
 691                 if (err)
 692                         extent_io_tree_panic(tree, err);
 693
 694                 prealloc = NULL;
 695                 if (err)
 696                         goto out;
 697                 if (state->end <= end) {
 698                         state = clear_state_bit(tree, state, &bits, wake,
 699                                                 changeset);
 700                         goto next;
 701                 }
 702                 goto search_again;
 703         }
 704         /*
 705          * | ---- desired range ---- |
 706          *                        | state |
 707          * We need to split the extent, and clear the bit
 708          * on the first half
 709          */
 710         if (state->start <= end && state->end > end) {
 711                 prealloc = alloc_extent_state_atomic(prealloc);
 712                 BUG_ON(!prealloc);
 713                 err = split_state(tree, state, prealloc, end + 1);
 714                 if (err)
 715                         extent_io_tree_panic(tree, err);
 716
 717                 if (wake)
 718                         wake_up(&state->wq);
 719
 720                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 721
 722                 prealloc = NULL;
 723                 goto out;
 724         }
 725
 726         state = clear_state_bit(tree, state, &bits, wake, changeset);
 727 next:
 728         if (last_end == (u64)-1)
 729                 goto out;
 730         start = last_end + 1;
 731         if (start <= end && state && !need_resched())
 732                 goto hit_next;
 733
 734 search_again:
 735         if (start > end)
 736                 goto out;
 737         spin_unlock(&tree->lock);
 738         if (gfpflags_allow_blocking(mask))
 739                 cond_resched();
 740         goto again;
 741
 742 out:
 743         spin_unlock(&tree->lock);
 744         if (prealloc)
 745                 free_extent_state(prealloc);
 746
 747         return 0;
 748
 749 }
 750
 751 static void wait_on_state(struct extent_io_tree *tree,
 752                           struct extent_state *state)
 753                 __releases(tree->lock)
 754                 __acquires(tree->lock)
 755 {
 756         DEFINE_WAIT(wait);
 757         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 758         spin_unlock(&tree->lock);
 759         schedule();
 760         spin_lock(&tree->lock);
 761         finish_wait(&state->wq, &wait);
 762 }
 763
 764 /*
 765  * waits for one or more bits to clear on a range in the state tree.
 766  * The range [start, end] is inclusive.
 767  * The tree lock is taken by this function
 768  */
 769 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 770                             unsigned long bits)
 771 {
 772         struct extent_state *state;
 773         struct rb_node *node;
 774
 775         btrfs_debug_check_extent_io_range(tree, start, end);
 776
 777         spin_lock(&tree->lock);
 778 again:
 779         while (1) {
 780                 /*
 781                  * this search will find all the extents that end after
 782                  * our range starts
 783                  */
 784                 node = tree_search(tree, start);
 785 process_node:
 786                 if (!node)
 787                         break;
 788
 789                 state = rb_entry(node, struct extent_state, rb_node);
 790
 791                 if (state->start > end)
 792                         goto out;
 793
 794                 if (state->state & bits) {
 795                         start = state->start;
 796                         atomic_inc(&state->refs);
 797                         wait_on_state(tree, state);
 798                         free_extent_state(state);
 799                         goto again;
 800                 }
 801                 start = state->end + 1;
 802
 803                 if (start > end)
 804                         break;
 805
 806                 if (!cond_resched_lock(&tree->lock)) {
 807                         node = rb_next(node);
 808                         goto process_node;
 809                 }
 810         }
 811 out:
 812         spin_unlock(&tree->lock);
 813 }
 814
 815 static void set_state_bits(struct extent_io_tree *tree,
 816                            struct extent_state *state,
 817                            unsigned *bits, struct extent_changeset *changeset)
 818 {
 819         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 820
 821         set_state_cb(tree, state, bits);
 822         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 823                 u64 range = state->end - state->start + 1;
 824                 tree->dirty_bytes += range;
 825         }
 826         add_extent_changeset(state, bits_to_set, changeset, 1);
 827         state->state |= bits_to_set;
 828 }
 829
 830 static void cache_state_if_flags(struct extent_state *state,
 831                                  struct extent_state **cached_ptr,
 832                                  unsigned flags)
 833 {
 834         if (cached_ptr && !(*cached_ptr)) {
 835                 if (!flags || (state->state & flags)) {
 836                         *cached_ptr = state;
 837                         atomic_inc(&state->refs);
 838                 }
 839         }
 840 }
 841
 842 static void cache_state(struct extent_state *state,
 843                         struct extent_state **cached_ptr)
 844 {
 845         return cache_state_if_flags(state, cached_ptr,
 846                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 847 }
 848
 849 /*
 850  * set some bits on a range in the tree.  This may require allocations or
 851  * sleeping, so the gfp mask is used to indicate what is allowed.
 852  *
 853  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 854  * part of the range already has the desired bits set.  The start of the
 855  * existing range is returned in failed_start in this case.
 856  *
 857  * [start, end] is inclusive This takes the tree lock.
 858  */
 859
 860 static int __must_check
 861 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 862                  unsigned bits, unsigned exclusive_bits,
 863                  u64 *failed_start, struct extent_state **cached_state,
 864                  gfp_t mask, struct extent_changeset *changeset)
 865 {
 866         struct extent_state *state;
 867         struct extent_state *prealloc = NULL;
 868         struct rb_node *node;
 869         struct rb_node **p;
 870         struct rb_node *parent;
 871         int err = 0;
 872         u64 last_start;
 873         u64 last_end;
 874
 875         btrfs_debug_check_extent_io_range(tree, start, end);
 876
 877         bits |= EXTENT_FIRST_DELALLOC;
 878 again:
 879         if (!prealloc && gfpflags_allow_blocking(mask)) {
 880                 /*
 881                  * Don't care for allocation failure here because we might end
 882                  * up not needing the pre-allocated extent state at all, which
 883                  * is the case if we only have in the tree extent states that
 884                  * cover our input range and don't cover too any other range.
 885                  * If we end up needing a new extent state we allocate it later.
 886                  */
 887                 prealloc = alloc_extent_state(mask);
 888         }
 889
 890         spin_lock(&tree->lock);
 891         if (cached_state && *cached_state) {
 892                 state = *cached_state;
 893                 if (state->start <= start && state->end > start &&
 894                     extent_state_in_tree(state)) {
 895                         node = &state->rb_node;
 896                         goto hit_next;
 897                 }
 898         }
 899         /*
 900          * this search will find all the extents that end after
 901          * our range starts.
 902          */
 903         node = tree_search_for_insert(tree, start, &p, &parent);
 904         if (!node) {
 905                 prealloc = alloc_extent_state_atomic(prealloc);
 906                 BUG_ON(!prealloc);
 907                 err = insert_state(tree, prealloc, start, end,
 908                                    &p, &parent, &bits, changeset);
 909                 if (err)
 910                         extent_io_tree_panic(tree, err);
 911
 912                 cache_state(prealloc, cached_state);
 913                 prealloc = NULL;
 914                 goto out;
 915         }
 916         state = rb_entry(node, struct extent_state, rb_node);
 917 hit_next:
 918         last_start = state->start;
 919         last_end = state->end;
 920
 921         /*
 922          * | ---- desired range ---- |
 923          * | state |
 924          *
 925          * Just lock what we found and keep going
 926          */
 927         if (state->start == start && state->end <= end) {
 928                 if (state->state & exclusive_bits) {
 929                         *failed_start = state->start;
 930                         err = -EEXIST;
 931                         goto out;
 932                 }
 933
 934                 set_state_bits(tree, state, &bits, changeset);
 935                 cache_state(state, cached_state);
 936                 merge_state(tree, state);
 937                 if (last_end == (u64)-1)
 938                         goto out;
 939                 start = last_end + 1;
 940                 state = next_state(state);
 941                 if (start < end && state && state->start == start &&
 942                     !need_resched())
 943                         goto hit_next;
 944                 goto search_again;
 945         }
 946
 947         /*
 948          *     | ---- desired range ---- |
 949          * | state |
 950          *   or
 951          * | ------------- state -------------- |
 952          *
 953          * We need to split the extent we found, and may flip bits on
 954          * second half.
 955          *
 956          * If the extent we found extends past our
 957          * range, we just split and search again.  It'll get split
 958          * again the next time though.
 959          *
 960          * If the extent we found is inside our range, we set the
 961          * desired bit on it.
 962          */
 963         if (state->start < start) {
 964                 if (state->state & exclusive_bits) {
 965                         *failed_start = start;
 966                         err = -EEXIST;
 967                         goto out;
 968                 }
 969
 970                 prealloc = alloc_extent_state_atomic(prealloc);
 971                 BUG_ON(!prealloc);
 972                 err = split_state(tree, state, prealloc, start);
 973                 if (err)
 974                         extent_io_tree_panic(tree, err);
 975
 976                 prealloc = NULL;
 977                 if (err)
 978                         goto out;
 979                 if (state->end <= end) {
 980                         set_state_bits(tree, state, &bits, changeset);
 981                         cache_state(state, cached_state);
 982                         merge_state(tree, state);
 983                         if (last_end == (u64)-1)
 984                                 goto out;
 985                         start = last_end + 1;
 986                         state = next_state(state);
 987                         if (start < end && state && state->start == start &&
 988                             !need_resched())
 989                                 goto hit_next;
 990                 }
 991                 goto search_again;
 992         }
 993         /*
 994          * | ---- desired range ---- |
 995          *     | state | or               | state |
 996          *
 997          * There's a hole, we need to insert something in it and
 998          * ignore the extent we found.
 999          */
1000         if (state->start > start) {
1001                 u64 this_end;
1002                 if (end < last_start)
1003                         this_end = end;
1004                 else
1005                         this_end = last_start - 1;
1006
1007                 prealloc = alloc_extent_state_atomic(prealloc);
1008                 BUG_ON(!prealloc);
1009
1010                 /*
1011                  * Avoid to free 'prealloc' if it can be merged with
1012                  * the later extent.
1013                  */
1014                 err = insert_state(tree, prealloc, start, this_end,
1015                                    NULL, NULL, &bits, changeset);
1016                 if (err)
1017                         extent_io_tree_panic(tree, err);
1018
1019                 cache_state(prealloc, cached_state);
1020                 prealloc = NULL;
1021                 start = this_end + 1;
1022                 goto search_again;
1023         }
1024         /*
1025          * | ---- desired range ---- |
1026          *                        | state |
1027          * We need to split the extent, and set the bit
1028          * on the first half
1029          */
1030         if (state->start <= end && state->end > end) {
1031                 if (state->state & exclusive_bits) {
1032                         *failed_start = start;
1033                         err = -EEXIST;
1034                         goto out;
1035                 }
1036
1037                 prealloc = alloc_extent_state_atomic(prealloc);
1038                 BUG_ON(!prealloc);
1039                 err = split_state(tree, state, prealloc, end + 1);
1040                 if (err)
1041                         extent_io_tree_panic(tree, err);
1042
1043                 set_state_bits(tree, prealloc, &bits, changeset);
1044                 cache_state(prealloc, cached_state);
1045                 merge_state(tree, prealloc);
1046                 prealloc = NULL;
1047                 goto out;
1048         }
1049
1050 search_again:
1051         if (start > end)
1052                 goto out;
1053         spin_unlock(&tree->lock);
1054         if (gfpflags_allow_blocking(mask))
1055                 cond_resched();
1056         goto again;
1057
1058 out:
1059         spin_unlock(&tree->lock);
1060         if (prealloc)
1061                 free_extent_state(prealloc);
1062
1063         return err;
1064
1065 }
1066
1067 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1068                    unsigned bits, u64 * failed_start,
1069                    struct extent_state **cached_state, gfp_t mask)
1070 {
1071         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1072                                 cached_state, mask, NULL);
1073 }
1074
1075
1076 /**
1077  * convert_extent_bit - convert all bits in a given range from one bit to
1078  *                      another
1079  * @tree:       the io tree to search
1080  * @start:      the start offset in bytes
1081  * @end:        the end offset in bytes (inclusive)
1082  * @bits:       the bits to set in this range
1083  * @clear_bits: the bits to clear in this range
1084  * @cached_state:       state that we're going to cache
1085  *
1086  * This will go through and set bits for the given range.  If any states exist
1087  * already in this range they are set with the given bit and cleared of the
1088  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1089  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1090  * boundary bits like LOCK.
1091  *
1092  * All allocations are done with GFP_NOFS.
1093  */
1094 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1095                        unsigned bits, unsigned clear_bits,
1096                        struct extent_state **cached_state)
1097 {
1098         struct extent_state *state;
1099         struct extent_state *prealloc = NULL;
1100         struct rb_node *node;
1101         struct rb_node **p;
1102         struct rb_node *parent;
1103         int err = 0;
1104         u64 last_start;
1105         u64 last_end;
1106         bool first_iteration = true;
1107
1108         btrfs_debug_check_extent_io_range(tree, start, end);
1109
1110 again:
1111         if (!prealloc) {
1112                 /*
1113                  * Best effort, don't worry if extent state allocation fails
1114                  * here for the first iteration. We might have a cached state
1115                  * that matches exactly the target range, in which case no
1116                  * extent state allocations are needed. We'll only know this
1117                  * after locking the tree.
1118                  */
1119                 prealloc = alloc_extent_state(GFP_NOFS);
1120                 if (!prealloc && !first_iteration)
1121                         return -ENOMEM;
1122         }
1123
1124         spin_lock(&tree->lock);
1125         if (cached_state && *cached_state) {
1126                 state = *cached_state;
1127                 if (state->start <= start && state->end > start &&
1128                     extent_state_in_tree(state)) {
1129                         node = &state->rb_node;
1130                         goto hit_next;
1131                 }
1132         }
1133
1134         /*
1135          * this search will find all the extents that end after
1136          * our range starts.
1137          */
1138         node = tree_search_for_insert(tree, start, &p, &parent);
1139         if (!node) {
1140                 prealloc = alloc_extent_state_atomic(prealloc);
1141                 if (!prealloc) {
1142                         err = -ENOMEM;
1143                         goto out;
1144                 }
1145                 err = insert_state(tree, prealloc, start, end,
1146                                    &p, &parent, &bits, NULL);
1147                 if (err)
1148                         extent_io_tree_panic(tree, err);
1149                 cache_state(prealloc, cached_state);
1150                 prealloc = NULL;
1151                 goto out;
1152         }
1153         state = rb_entry(node, struct extent_state, rb_node);
1154 hit_next:
1155         last_start = state->start;
1156         last_end = state->end;
1157
1158         /*
1159          * | ---- desired range ---- |
1160          * | state |
1161          *
1162          * Just lock what we found and keep going
1163          */
1164         if (state->start == start && state->end <= end) {
1165                 set_state_bits(tree, state, &bits, NULL);
1166                 cache_state(state, cached_state);
1167                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1168                 if (last_end == (u64)-1)
1169                         goto out;
1170                 start = last_end + 1;
1171                 if (start < end && state && state->start == start &&
1172                     !need_resched())
1173                         goto hit_next;
1174                 goto search_again;
1175         }
1176
1177         /*
1178          *     | ---- desired range ---- |
1179          * | state |
1180          *   or
1181          * | ------------- state -------------- |
1182          *
1183          * We need to split the extent we found, and may flip bits on
1184          * second half.
1185          *
1186          * If the extent we found extends past our
1187          * range, we just split and search again.  It'll get split
1188          * again the next time though.
1189          *
1190          * If the extent we found is inside our range, we set the
1191          * desired bit on it.
1192          */
1193         if (state->start < start) {
1194                 prealloc = alloc_extent_state_atomic(prealloc);
1195                 if (!prealloc) {
1196                         err = -ENOMEM;
1197                         goto out;
1198                 }
1199                 err = split_state(tree, state, prealloc, start);
1200                 if (err)
1201                         extent_io_tree_panic(tree, err);
1202                 prealloc = NULL;
1203                 if (err)
1204                         goto out;
1205                 if (state->end <= end) {
1206                         set_state_bits(tree, state, &bits, NULL);
1207                         cache_state(state, cached_state);
1208                         state = clear_state_bit(tree, state, &clear_bits, 0,
1209                                                 NULL);
1210                         if (last_end == (u64)-1)
1211                                 goto out;
1212                         start = last_end + 1;
1213                         if (start < end && state && state->start == start &&
1214                             !need_resched())
1215                                 goto hit_next;
1216                 }
1217                 goto search_again;
1218         }
1219         /*
1220          * | ---- desired range ---- |
1221          *     | state | or               | state |
1222          *
1223          * There's a hole, we need to insert something in it and
1224          * ignore the extent we found.
1225          */
1226         if (state->start > start) {
1227                 u64 this_end;
1228                 if (end < last_start)
1229                         this_end = end;
1230                 else
1231                         this_end = last_start - 1;
1232
1233                 prealloc = alloc_extent_state_atomic(prealloc);
1234                 if (!prealloc) {
1235                         err = -ENOMEM;
1236                         goto out;
1237                 }
1238
1239                 /*
1240                  * Avoid to free 'prealloc' if it can be merged with
1241                  * the later extent.
1242                  */
1243                 err = insert_state(tree, prealloc, start, this_end,
1244                                    NULL, NULL, &bits, NULL);
1245                 if (err)
1246                         extent_io_tree_panic(tree, err);
1247                 cache_state(prealloc, cached_state);
1248                 prealloc = NULL;
1249                 start = this_end + 1;
1250                 goto search_again;
1251         }
1252         /*
1253          * | ---- desired range ---- |
1254          *                        | state |
1255          * We need to split the extent, and set the bit
1256          * on the first half
1257          */
1258         if (state->start <= end && state->end > end) {
1259                 prealloc = alloc_extent_state_atomic(prealloc);
1260                 if (!prealloc) {
1261                         err = -ENOMEM;
1262                         goto out;
1263                 }
1264
1265                 err = split_state(tree, state, prealloc, end + 1);
1266                 if (err)
1267                         extent_io_tree_panic(tree, err);
1268
1269                 set_state_bits(tree, prealloc, &bits, NULL);
1270                 cache_state(prealloc, cached_state);
1271                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1272                 prealloc = NULL;
1273                 goto out;
1274         }
1275
1276 search_again:
1277         if (start > end)
1278                 goto out;
1279         spin_unlock(&tree->lock);
1280         cond_resched();
1281         first_iteration = false;
1282         goto again;
1283
1284 out:
1285         spin_unlock(&tree->lock);
1286         if (prealloc)
1287                 free_extent_state(prealloc);
1288
1289         return err;
1290 }
1291
1292 /* wrappers around set/clear extent bit */
1293 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1294                            unsigned bits, struct extent_changeset *changeset)
1295 {
1296         /*
1297          * We don't support EXTENT_LOCKED yet, as current changeset will
1298          * record any bits changed, so for EXTENT_LOCKED case, it will
1299          * either fail with -EEXIST or changeset will record the whole
1300          * range.
1301          */
1302         BUG_ON(bits & EXTENT_LOCKED);
1303
1304         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1305                                 changeset);
1306 }
1307
1308 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1309                      unsigned bits, int wake, int delete,
1310                      struct extent_state **cached, gfp_t mask)
1311 {
1312         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1313                                   cached, mask, NULL);
1314 }
1315
1316 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1317                 unsigned bits, struct extent_changeset *changeset)
1318 {
1319         /*
1320          * Don't support EXTENT_LOCKED case, same reason as
1321          * set_record_extent_bits().
1322          */
1323         BUG_ON(bits & EXTENT_LOCKED);
1324
1325         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1326                                   changeset);
1327 }
1328
1329 /*
1330  * either insert or lock state struct between start and end use mask to tell
1331  * us if waiting is desired.
1332  */
1333 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1334                      struct extent_state **cached_state)
1335 {
1336         int err;
1337         u64 failed_start;
1338
1339         while (1) {
1340                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1341                                        EXTENT_LOCKED, &failed_start,
1342                                        cached_state, GFP_NOFS, NULL);
1343                 if (err == -EEXIST) {
1344                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1345                         start = failed_start;
1346                 } else
1347                         break;
1348                 WARN_ON(start > end);
1349         }
1350         return err;
1351 }
1352
1353 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1354 {
1355         int err;
1356         u64 failed_start;
1357
1358         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1359                                &failed_start, NULL, GFP_NOFS, NULL);
1360         if (err == -EEXIST) {
1361                 if (failed_start > start)
1362                         clear_extent_bit(tree, start, failed_start - 1,
1363                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1364                 return 0;
1365         }
1366         return 1;
1367 }
1368
1369 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1370 {
1371         unsigned long index = start >> PAGE_SHIFT;
1372         unsigned long end_index = end >> PAGE_SHIFT;
1373         struct page *page;
1374
1375         while (index <= end_index) {
1376                 page = find_get_page(inode->i_mapping, index);
1377                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1378                 clear_page_dirty_for_io(page);
1379                 put_page(page);
1380                 index++;
1381         }
1382 }
1383
1384 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1385 {
1386         unsigned long index = start >> PAGE_SHIFT;
1387         unsigned long end_index = end >> PAGE_SHIFT;
1388         struct page *page;
1389
1390         while (index <= end_index) {
1391                 page = find_get_page(inode->i_mapping, index);
1392                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1393                 __set_page_dirty_nobuffers(page);
1394                 account_page_redirty(page);
1395                 put_page(page);
1396                 index++;
1397         }
1398 }
1399
1400 /*
1401  * helper function to set both pages and extents in the tree writeback
1402  */
1403 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1404 {
1405         unsigned long index = start >> PAGE_SHIFT;
1406         unsigned long end_index = end >> PAGE_SHIFT;
1407         struct page *page;
1408
1409         while (index <= end_index) {
1410                 page = find_get_page(tree->mapping, index);
1411                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1412                 set_page_writeback(page);
1413                 put_page(page);
1414                 index++;
1415         }
1416 }
1417
1418 /* find the first state struct with 'bits' set after 'start', and
1419  * return it.  tree->lock must be held.  NULL will returned if
1420  * nothing was found after 'start'
1421  */
1422 static struct extent_state *
1423 find_first_extent_bit_state(struct extent_io_tree *tree,
1424                             u64 start, unsigned bits)
1425 {
1426         struct rb_node *node;
1427         struct extent_state *state;
1428
1429         /*
1430          * this search will find all the extents that end after
1431          * our range starts.
1432          */
1433         node = tree_search(tree, start);
1434         if (!node)
1435                 goto out;
1436
1437         while (1) {
1438                 state = rb_entry(node, struct extent_state, rb_node);
1439                 if (state->end >= start && (state->state & bits))
1440                         return state;
1441
1442                 node = rb_next(node);
1443                 if (!node)
1444                         break;
1445         }
1446 out:
1447         return NULL;
1448 }
1449
1450 /*
1451  * find the first offset in the io tree with 'bits' set. zero is
1452  * returned if we find something, and *start_ret and *end_ret are
1453  * set to reflect the state struct that was found.
1454  *
1455  * If nothing was found, 1 is returned. If found something, return 0.
1456  */
1457 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1458                           u64 *start_ret, u64 *end_ret, unsigned bits,
1459                           struct extent_state **cached_state)
1460 {
1461         struct extent_state *state;
1462         struct rb_node *n;
1463         int ret = 1;
1464
1465         spin_lock(&tree->lock);
1466         if (cached_state && *cached_state) {
1467                 state = *cached_state;
1468                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1469                         n = rb_next(&state->rb_node);
1470                         while (n) {
1471                                 state = rb_entry(n, struct extent_state,
1472                                                  rb_node);
1473                                 if (state->state & bits)
1474                                         goto got_it;
1475                                 n = rb_next(n);
1476                         }
1477                         free_extent_state(*cached_state);
1478                         *cached_state = NULL;
1479                         goto out;
1480                 }
1481                 free_extent_state(*cached_state);
1482                 *cached_state = NULL;
1483         }
1484
1485         state = find_first_extent_bit_state(tree, start, bits);
1486 got_it:
1487         if (state) {
1488                 cache_state_if_flags(state, cached_state, 0);
1489                 *start_ret = state->start;
1490                 *end_ret = state->end;
1491                 ret = 0;
1492         }
1493 out:
1494         spin_unlock(&tree->lock);
1495         return ret;
1496 }
1497
1498 /*
1499  * find a contiguous range of bytes in the file marked as delalloc, not
1500  * more than 'max_bytes'.  start and end are used to return the range,
1501  *
1502  * 1 is returned if we find something, 0 if nothing was in the tree
1503  */
1504 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1505                                         u64 *start, u64 *end, u64 max_bytes,
1506                                         struct extent_state **cached_state)
1507 {
1508         struct rb_node *node;
1509         struct extent_state *state;
1510         u64 cur_start = *start;
1511         u64 found = 0;
1512         u64 total_bytes = 0;
1513
1514         spin_lock(&tree->lock);
1515
1516         /*
1517          * this search will find all the extents that end after
1518          * our range starts.
1519          */
1520         node = tree_search(tree, cur_start);
1521         if (!node) {
1522                 if (!found)
1523                         *end = (u64)-1;
1524                 goto out;
1525         }
1526
1527         while (1) {
1528                 state = rb_entry(node, struct extent_state, rb_node);
1529                 if (found && (state->start != cur_start ||
1530                               (state->state & EXTENT_BOUNDARY))) {
1531                         goto out;
1532                 }
1533                 if (!(state->state & EXTENT_DELALLOC)) {
1534                         if (!found)
1535                                 *end = state->end;
1536                         goto out;
1537                 }
1538                 if (!found) {
1539                         *start = state->start;
1540                         *cached_state = state;
1541                         atomic_inc(&state->refs);
1542                 }
1543                 found++;
1544                 *end = state->end;
1545                 cur_start = state->end + 1;
1546                 node = rb_next(node);
1547                 total_bytes += state->end - state->start + 1;
1548                 if (total_bytes >= max_bytes)
1549                         break;
1550                 if (!node)
1551                         break;
1552         }
1553 out:
1554         spin_unlock(&tree->lock);
1555         return found;
1556 }
1557
1558 static int __process_pages_contig(struct address_space *mapping,
1559                                   struct page *locked_page,
1560                                   pgoff_t start_index, pgoff_t end_index,
1561                                   unsigned long page_ops, pgoff_t *index_ret);
1562
1563 static noinline void __unlock_for_delalloc(struct inode *inode,
1564                                            struct page *locked_page,
1565                                            u64 start, u64 end)
1566 {
1567         unsigned long index = start >> PAGE_SHIFT;
1568         unsigned long end_index = end >> PAGE_SHIFT;
1569
1570         ASSERT(locked_page);
1571         if (index == locked_page->index && end_index == index)
1572                 return;
1573
1574         __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1575                                PAGE_UNLOCK, NULL);
1576 }
1577
1578 static noinline int lock_delalloc_pages(struct inode *inode,
1579                                         struct page *locked_page,
1580                                         u64 delalloc_start,
1581                                         u64 delalloc_end)
1582 {
1583         unsigned long index = delalloc_start >> PAGE_SHIFT;
1584         unsigned long index_ret = index;
1585         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1586         int ret;
1587
1588         ASSERT(locked_page);
1589         if (index == locked_page->index && index == end_index)
1590                 return 0;
1591
1592         ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1593                                      end_index, PAGE_LOCK, &index_ret);
1594         if (ret == -EAGAIN)
1595                 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1596                                       (u64)index_ret << PAGE_SHIFT);
1597         return ret;
1598 }
1599
1600 /*
1601  * find a contiguous range of bytes in the file marked as delalloc, not
1602  * more than 'max_bytes'.  start and end are used to return the range,
1603  *
1604  * 1 is returned if we find something, 0 if nothing was in the tree
1605  */
1606 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1607                                     struct extent_io_tree *tree,
1608                                     struct page *locked_page, u64 *start,
1609                                     u64 *end, u64 max_bytes)
1610 {
1611         u64 delalloc_start;
1612         u64 delalloc_end;
1613         u64 found;
1614         struct extent_state *cached_state = NULL;
1615         int ret;
1616         int loops = 0;
1617
1618 again:
1619         /* step one, find a bunch of delalloc bytes starting at start */
1620         delalloc_start = *start;
1621         delalloc_end = 0;
1622         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1623                                     max_bytes, &cached_state);
1624         if (!found || delalloc_end <= *start) {
1625                 *start = delalloc_start;
1626                 *end = delalloc_end;
1627                 free_extent_state(cached_state);
1628                 return 0;
1629         }
1630
1631         /*
1632          * start comes from the offset of locked_page.  We have to lock
1633          * pages in order, so we can't process delalloc bytes before
1634          * locked_page
1635          */
1636         if (delalloc_start < *start)
1637                 delalloc_start = *start;
1638
1639         /*
1640          * make sure to limit the number of pages we try to lock down
1641          */
1642         if (delalloc_end + 1 - delalloc_start > max_bytes)
1643                 delalloc_end = delalloc_start + max_bytes - 1;
1644
1645         /* step two, lock all the pages after the page that has start */
1646         ret = lock_delalloc_pages(inode, locked_page,
1647                                   delalloc_start, delalloc_end);
1648         if (ret == -EAGAIN) {
1649                 /* some of the pages are gone, lets avoid looping by
1650                  * shortening the size of the delalloc range we're searching
1651                  */
1652                 free_extent_state(cached_state);
1653                 cached_state = NULL;
1654                 if (!loops) {
1655                         max_bytes = PAGE_SIZE;
1656                         loops = 1;
1657                         goto again;
1658                 } else {
1659                         found = 0;
1660                         goto out_failed;
1661                 }
1662         }
1663         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1664
1665         /* step three, lock the state bits for the whole range */
1666         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1667
1668         /* then test to make sure it is all still delalloc */
1669         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1670                              EXTENT_DELALLOC, 1, cached_state);
1671         if (!ret) {
1672                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1673                                      &cached_state, GFP_NOFS);
1674                 __unlock_for_delalloc(inode, locked_page,
1675                               delalloc_start, delalloc_end);
1676                 cond_resched();
1677                 goto again;
1678         }
1679         free_extent_state(cached_state);
1680         *start = delalloc_start;
1681         *end = delalloc_end;
1682 out_failed:
1683         return found;
1684 }
1685
1686 static int __process_pages_contig(struct address_space *mapping,
1687                                   struct page *locked_page,
1688                                   pgoff_t start_index, pgoff_t end_index,
1689                                   unsigned long page_ops, pgoff_t *index_ret)
1690 {
1691         unsigned long nr_pages = end_index - start_index + 1;
1692         unsigned long pages_locked = 0;
1693         pgoff_t index = start_index;
1694         struct page *pages[16];
1695         unsigned ret;
1696         int err = 0;
1697         int i;
1698
1699         if (page_ops & PAGE_LOCK) {
1700                 ASSERT(page_ops == PAGE_LOCK);
1701                 ASSERT(index_ret && *index_ret == start_index);
1702         }
1703
1704         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1705                 mapping_set_error(mapping, -EIO);
1706
1707         while (nr_pages > 0) {
1708                 ret = find_get_pages_contig(mapping, index,
1709                                      min_t(unsigned long,
1710                                      nr_pages, ARRAY_SIZE(pages)), pages);
1711                 if (ret == 0) {
1712                         /*
1713                          * Only if we're going to lock these pages,
1714                          * can we find nothing at @index.
1715                          */
1716                         ASSERT(page_ops & PAGE_LOCK);
1717                         return ret;
1718                 }
1719
1720                 for (i = 0; i < ret; i++) {
1721                         if (page_ops & PAGE_SET_PRIVATE2)
1722                                 SetPagePrivate2(pages[i]);
1723
1724                         if (pages[i] == locked_page) {
1725                                 put_page(pages[i]);
1726                                 pages_locked++;
1727                                 continue;
1728                         }
1729                         if (page_ops & PAGE_CLEAR_DIRTY)
1730                                 clear_page_dirty_for_io(pages[i]);
1731                         if (page_ops & PAGE_SET_WRITEBACK)
1732                                 set_page_writeback(pages[i]);
1733                         if (page_ops & PAGE_SET_ERROR)
1734                                 SetPageError(pages[i]);
1735                         if (page_ops & PAGE_END_WRITEBACK)
1736                                 end_page_writeback(pages[i]);
1737                         if (page_ops & PAGE_UNLOCK)
1738                                 unlock_page(pages[i]);
1739                         if (page_ops & PAGE_LOCK) {
1740                                 lock_page(pages[i]);
1741                                 if (!PageDirty(pages[i]) ||
1742                                     pages[i]->mapping != mapping) {
1743                                         unlock_page(pages[i]);
1744                                         put_page(pages[i]);
1745                                         err = -EAGAIN;
1746                                         goto out;
1747                                 }
1748                         }
1749                         put_page(pages[i]);
1750                         pages_locked++;
1751                 }
1752                 nr_pages -= ret;
1753                 index += ret;
1754                 cond_resched();
1755         }
1756 out:
1757         if (err && index_ret)
1758                 *index_ret = start_index + pages_locked - 1;
1759         return err;
1760 }
1761
1762 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1763                                  u64 delalloc_end, struct page *locked_page,
1764                                  unsigned clear_bits,
1765                                  unsigned long page_ops)
1766 {
1767         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1768                          NULL, GFP_NOFS);
1769
1770         __process_pages_contig(inode->i_mapping, locked_page,
1771                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1772                                page_ops, NULL);
1773 }
1774
1775 /*
1776  * count the number of bytes in the tree that have a given bit(s)
1777  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1778  * cached.  The total number found is returned.
1779  */
1780 u64 count_range_bits(struct extent_io_tree *tree,
1781                      u64 *start, u64 search_end, u64 max_bytes,
1782                      unsigned bits, int contig)
1783 {
1784         struct rb_node *node;
1785         struct extent_state *state;
1786         u64 cur_start = *start;
1787         u64 total_bytes = 0;
1788         u64 last = 0;
1789         int found = 0;
1790
1791         if (WARN_ON(search_end <= cur_start))
1792                 return 0;
1793
1794         spin_lock(&tree->lock);
1795         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1796                 total_bytes = tree->dirty_bytes;
1797                 goto out;
1798         }
1799         /*
1800          * this search will find all the extents that end after
1801          * our range starts.
1802          */
1803         node = tree_search(tree, cur_start);
1804         if (!node)
1805                 goto out;
1806
1807         while (1) {
1808                 state = rb_entry(node, struct extent_state, rb_node);
1809                 if (state->start > search_end)
1810                         break;
1811                 if (contig && found && state->start > last + 1)
1812                         break;
1813                 if (state->end >= cur_start && (state->state & bits) == bits) {
1814                         total_bytes += min(search_end, state->end) + 1 -
1815                                        max(cur_start, state->start);
1816                         if (total_bytes >= max_bytes)
1817                                 break;
1818                         if (!found) {
1819                                 *start = max(cur_start, state->start);
1820                                 found = 1;
1821                         }
1822                         last = state->end;
1823                 } else if (contig && found) {
1824                         break;
1825                 }
1826                 node = rb_next(node);
1827                 if (!node)
1828                         break;
1829         }
1830 out:
1831         spin_unlock(&tree->lock);
1832         return total_bytes;
1833 }
1834
1835 /*
1836  * set the private field for a given byte offset in the tree.  If there isn't
1837  * an extent_state there already, this does nothing.
1838  */
1839 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1840                 struct io_failure_record *failrec)
1841 {
1842         struct rb_node *node;
1843         struct extent_state *state;
1844         int ret = 0;
1845
1846         spin_lock(&tree->lock);
1847         /*
1848          * this search will find all the extents that end after
1849          * our range starts.
1850          */
1851         node = tree_search(tree, start);
1852         if (!node) {
1853                 ret = -ENOENT;
1854                 goto out;
1855         }
1856         state = rb_entry(node, struct extent_state, rb_node);
1857         if (state->start != start) {
1858                 ret = -ENOENT;
1859                 goto out;
1860         }
1861         state->failrec = failrec;
1862 out:
1863         spin_unlock(&tree->lock);
1864         return ret;
1865 }
1866
1867 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1868                 struct io_failure_record **failrec)
1869 {
1870         struct rb_node *node;
1871         struct extent_state *state;
1872         int ret = 0;
1873
1874         spin_lock(&tree->lock);
1875         /*
1876          * this search will find all the extents that end after
1877          * our range starts.
1878          */
1879         node = tree_search(tree, start);
1880         if (!node) {
1881                 ret = -ENOENT;
1882                 goto out;
1883         }
1884         state = rb_entry(node, struct extent_state, rb_node);
1885         if (state->start != start) {
1886                 ret = -ENOENT;
1887                 goto out;
1888         }
1889         *failrec = state->failrec;
1890 out:
1891         spin_unlock(&tree->lock);
1892         return ret;
1893 }
1894
1895 /*
1896  * searches a range in the state tree for a given mask.
1897  * If 'filled' == 1, this returns 1 only if every extent in the tree
1898  * has the bits set.  Otherwise, 1 is returned if any bit in the
1899  * range is found set.
1900  */
1901 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1902                    unsigned bits, int filled, struct extent_state *cached)
1903 {
1904         struct extent_state *state = NULL;
1905         struct rb_node *node;
1906         int bitset = 0;
1907
1908         spin_lock(&tree->lock);
1909         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1910             cached->end > start)
1911                 node = &cached->rb_node;
1912         else
1913                 node = tree_search(tree, start);
1914         while (node && start <= end) {
1915                 state = rb_entry(node, struct extent_state, rb_node);
1916
1917                 if (filled && state->start > start) {
1918                         bitset = 0;
1919                         break;
1920                 }
1921
1922                 if (state->start > end)
1923                         break;
1924
1925                 if (state->state & bits) {
1926                         bitset = 1;
1927                         if (!filled)
1928                                 break;
1929                 } else if (filled) {
1930                         bitset = 0;
1931                         break;
1932                 }
1933
1934                 if (state->end == (u64)-1)
1935                         break;
1936
1937                 start = state->end + 1;
1938                 if (start > end)
1939                         break;
1940                 node = rb_next(node);
1941                 if (!node) {
1942                         if (filled)
1943                                 bitset = 0;
1944                         break;
1945                 }
1946         }
1947         spin_unlock(&tree->lock);
1948         return bitset;
1949 }
1950
1951 /*
1952  * helper function to set a given page up to date if all the
1953  * extents in the tree for that page are up to date
1954  */
1955 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1956 {
1957         u64 start = page_offset(page);
1958         u64 end = start + PAGE_SIZE - 1;
1959         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1960                 SetPageUptodate(page);
1961 }
1962
1963 int free_io_failure(struct btrfs_inode *inode, struct io_failure_record *rec)
1964 {
1965         int ret;
1966         int err = 0;
1967         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
1968
1969         set_state_failrec(failure_tree, rec->start, NULL);
1970         ret = clear_extent_bits(failure_tree, rec->start,
1971                                 rec->start + rec->len - 1,
1972                                 EXTENT_LOCKED | EXTENT_DIRTY);
1973         if (ret)
1974                 err = ret;
1975
1976         ret = clear_extent_bits(&inode->io_tree, rec->start,
1977                                 rec->start + rec->len - 1,
1978                                 EXTENT_DAMAGED);
1979         if (ret && !err)
1980                 err = ret;
1981
1982         kfree(rec);
1983         return err;
1984 }
1985
1986 /*
1987  * this bypasses the standard btrfs submit functions deliberately, as
1988  * the standard behavior is to write all copies in a raid setup. here we only
1989  * want to write the one bad copy. so we do the mapping for ourselves and issue
1990  * submit_bio directly.
1991  * to avoid any synchronization issues, wait for the data after writing, which
1992  * actually prevents the read that triggered the error from finishing.
1993  * currently, there can be no more than two copies of every data bit. thus,
1994  * exactly one rewrite is required.
1995  */
1996 int repair_io_failure(struct btrfs_inode *inode, u64 start, u64 length,
1997                 u64 logical, struct page *page,
1998                 unsigned int pg_offset, int mirror_num)
1999 {
2000         struct btrfs_fs_info *fs_info = inode->root->fs_info;
2001         struct bio *bio;
2002         struct btrfs_device *dev;
2003         u64 map_length = 0;
2004         u64 sector;
2005         struct btrfs_bio *bbio = NULL;
2006         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2007         int ret;
2008
2009         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2010         BUG_ON(!mirror_num);
2011
2012         /* we can't repair anything in raid56 yet */
2013         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2014                 return 0;
2015
2016         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2017         if (!bio)
2018                 return -EIO;
2019         bio->bi_iter.bi_size = 0;
2020         map_length = length;
2021
2022         /*
2023          * Avoid races with device replace and make sure our bbio has devices
2024          * associated to its stripes that don't go away while we are doing the
2025          * read repair operation.
2026          */
2027         btrfs_bio_counter_inc_blocked(fs_info);
2028         ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2029                               &map_length, &bbio, mirror_num);
2030         if (ret) {
2031                 btrfs_bio_counter_dec(fs_info);
2032                 bio_put(bio);
2033                 return -EIO;
2034         }
2035         BUG_ON(mirror_num != bbio->mirror_num);
2036         sector = bbio->stripes[mirror_num-1].physical >> 9;
2037         bio->bi_iter.bi_sector = sector;
2038         dev = bbio->stripes[mirror_num-1].dev;
2039         btrfs_put_bbio(bbio);
2040         if (!dev || !dev->bdev || !dev->writeable) {
2041                 btrfs_bio_counter_dec(fs_info);
2042                 bio_put(bio);
2043                 return -EIO;
2044         }
2045         bio->bi_bdev = dev->bdev;
2046         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2047         bio_add_page(bio, page, length, pg_offset);
2048
2049         if (btrfsic_submit_bio_wait(bio)) {
2050                 /* try to remap that extent elsewhere? */
2051                 btrfs_bio_counter_dec(fs_info);
2052                 bio_put(bio);
2053                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2054                 return -EIO;
2055         }
2056
2057         btrfs_info_rl_in_rcu(fs_info,
2058                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2059                                   btrfs_ino(inode), start,
2060                                   rcu_str_deref(dev->name), sector);
2061         btrfs_bio_counter_dec(fs_info);
2062         bio_put(bio);
2063         return 0;
2064 }
2065
2066 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2067                          struct extent_buffer *eb, int mirror_num)
2068 {
2069         u64 start = eb->start;
2070         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2071         int ret = 0;
2072
2073         if (fs_info->sb->s_flags & MS_RDONLY)
2074                 return -EROFS;
2075
2076         for (i = 0; i < num_pages; i++) {
2077                 struct page *p = eb->pages[i];
2078
2079                 ret = repair_io_failure(BTRFS_I(fs_info->btree_inode), start,
2080                                         PAGE_SIZE, start, p,
2081                                         start - page_offset(p), mirror_num);
2082                 if (ret)
2083                         break;
2084                 start += PAGE_SIZE;
2085         }
2086
2087         return ret;
2088 }
2089
2090 /*
2091  * each time an IO finishes, we do a fast check in the IO failure tree
2092  * to see if we need to process or clean up an io_failure_record
2093  */
2094 int clean_io_failure(struct btrfs_inode *inode, u64 start, struct page *page,
2095                      unsigned int pg_offset)
2096 {
2097         u64 private;
2098         struct io_failure_record *failrec;
2099         struct btrfs_fs_info *fs_info = inode->root->fs_info;
2100         struct extent_state *state;
2101         int num_copies;
2102         int ret;
2103
2104         private = 0;
2105         ret = count_range_bits(&inode->io_failure_tree, &private,
2106                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2107         if (!ret)
2108                 return 0;
2109
2110         ret = get_state_failrec(&inode->io_failure_tree, start,
2111                         &failrec);
2112         if (ret)
2113                 return 0;
2114
2115         BUG_ON(!failrec->this_mirror);
2116
2117         if (failrec->in_validation) {
2118                 /* there was no real error, just free the record */
2119                 btrfs_debug(fs_info,
2120                         "clean_io_failure: freeing dummy error at %llu",
2121                         failrec->start);
2122                 goto out;
2123         }
2124         if (fs_info->sb->s_flags & MS_RDONLY)
2125                 goto out;
2126
2127         spin_lock(&inode->io_tree.lock);
2128         state = find_first_extent_bit_state(&inode->io_tree,
2129                                             failrec->start,
2130                                             EXTENT_LOCKED);
2131         spin_unlock(&inode->io_tree.lock);
2132
2133         if (state && state->start <= failrec->start &&
2134             state->end >= failrec->start + failrec->len - 1) {
2135                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2136                                               failrec->len);
2137                 if (num_copies > 1)  {
2138                         repair_io_failure(inode, start, failrec->len,
2139                                           failrec->logical, page,
2140                                           pg_offset, failrec->failed_mirror);
2141                 }
2142         }
2143
2144 out:
2145         free_io_failure(inode, failrec);
2146
2147         return 0;
2148 }
2149
2150 /*
2151  * Can be called when
2152  * - hold extent lock
2153  * - under ordered extent
2154  * - the inode is freeing
2155  */
2156 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2157 {
2158         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2159         struct io_failure_record *failrec;
2160         struct extent_state *state, *next;
2161
2162         if (RB_EMPTY_ROOT(&failure_tree->state))
2163                 return;
2164
2165         spin_lock(&failure_tree->lock);
2166         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2167         while (state) {
2168                 if (state->start > end)
2169                         break;
2170
2171                 ASSERT(state->end <= end);
2172
2173                 next = next_state(state);
2174
2175                 failrec = state->failrec;
2176                 free_extent_state(state);
2177                 kfree(failrec);
2178
2179                 state = next;
2180         }
2181         spin_unlock(&failure_tree->lock);
2182 }
2183
2184 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2185                 struct io_failure_record **failrec_ret)
2186 {
2187         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2188         struct io_failure_record *failrec;
2189         struct extent_map *em;
2190         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2191         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2192         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2193         int ret;
2194         u64 logical;
2195
2196         ret = get_state_failrec(failure_tree, start, &failrec);
2197         if (ret) {
2198                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2199                 if (!failrec)
2200                         return -ENOMEM;
2201
2202                 failrec->start = start;
2203                 failrec->len = end - start + 1;
2204                 failrec->this_mirror = 0;
2205                 failrec->bio_flags = 0;
2206                 failrec->in_validation = 0;
2207
2208                 read_lock(&em_tree->lock);
2209                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2210                 if (!em) {
2211                         read_unlock(&em_tree->lock);
2212                         kfree(failrec);
2213                         return -EIO;
2214                 }
2215
2216                 if (em->start > start || em->start + em->len <= start) {
2217                         free_extent_map(em);
2218                         em = NULL;
2219                 }
2220                 read_unlock(&em_tree->lock);
2221                 if (!em) {
2222                         kfree(failrec);
2223                         return -EIO;
2224                 }
2225
2226                 logical = start - em->start;
2227                 logical = em->block_start + logical;
2228                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2229                         logical = em->block_start;
2230                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2231                         extent_set_compress_type(&failrec->bio_flags,
2232                                                  em->compress_type);
2233                 }
2234
2235                 btrfs_debug(fs_info,
2236                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2237                         logical, start, failrec->len);
2238
2239                 failrec->logical = logical;
2240                 free_extent_map(em);
2241
2242                 /* set the bits in the private failure tree */
2243                 ret = set_extent_bits(failure_tree, start, end,
2244                                         EXTENT_LOCKED | EXTENT_DIRTY);
2245                 if (ret >= 0)
2246                         ret = set_state_failrec(failure_tree, start, failrec);
2247                 /* set the bits in the inode's tree */
2248                 if (ret >= 0)
2249                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2250                 if (ret < 0) {
2251                         kfree(failrec);
2252                         return ret;
2253                 }
2254         } else {
2255                 btrfs_debug(fs_info,
2256                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
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         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2275         int num_copies;
2276
2277         num_copies = btrfs_num_copies(fs_info, 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                 btrfs_debug(fs_info,
2285                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2286                         num_copies, failrec->this_mirror, failed_mirror);
2287                 return 0;
2288         }
2289
2290         /*
2291          * there are two premises:
2292          *      a) deliver good data to the caller
2293          *      b) correct the bad sectors on disk
2294          */
2295         if (failed_bio->bi_vcnt > 1) {
2296                 /*
2297                  * to fulfill b), we need to know the exact failing sectors, as
2298                  * we don't want to rewrite any more than the failed ones. thus,
2299                  * we need separate read requests for the failed bio
2300                  *
2301                  * if the following BUG_ON triggers, our validation request got
2302                  * merged. we need separate requests for our algorithm to work.
2303                  */
2304                 BUG_ON(failrec->in_validation);
2305                 failrec->in_validation = 1;
2306                 failrec->this_mirror = failed_mirror;
2307         } else {
2308                 /*
2309                  * we're ready to fulfill a) and b) alongside. get a good copy
2310                  * of the failed sector and if we succeed, we have setup
2311                  * everything for repair_io_failure to do the rest for us.
2312                  */
2313                 if (failrec->in_validation) {
2314                         BUG_ON(failrec->this_mirror != failed_mirror);
2315                         failrec->in_validation = 0;
2316                         failrec->this_mirror = 0;
2317                 }
2318                 failrec->failed_mirror = failed_mirror;
2319                 failrec->this_mirror++;
2320                 if (failrec->this_mirror == failed_mirror)
2321                         failrec->this_mirror++;
2322         }
2323
2324         if (failrec->this_mirror > num_copies) {
2325                 btrfs_debug(fs_info,
2326                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2327                         num_copies, failrec->this_mirror, failed_mirror);
2328                 return 0;
2329         }
2330
2331         return 1;
2332 }
2333
2334
2335 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2336                                     struct io_failure_record *failrec,
2337                                     struct page *page, int pg_offset, int icsum,
2338                                     bio_end_io_t *endio_func, void *data)
2339 {
2340         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2341         struct bio *bio;
2342         struct btrfs_io_bio *btrfs_failed_bio;
2343         struct btrfs_io_bio *btrfs_bio;
2344
2345         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2346         if (!bio)
2347                 return NULL;
2348
2349         bio->bi_end_io = endio_func;
2350         bio->bi_iter.bi_sector = failrec->logical >> 9;
2351         bio->bi_bdev = fs_info->fs_devices->latest_bdev;
2352         bio->bi_iter.bi_size = 0;
2353         bio->bi_private = data;
2354
2355         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2356         if (btrfs_failed_bio->csum) {
2357                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2358
2359                 btrfs_bio = btrfs_io_bio(bio);
2360                 btrfs_bio->csum = btrfs_bio->csum_inline;
2361                 icsum *= csum_size;
2362                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2363                        csum_size);
2364         }
2365
2366         bio_add_page(bio, page, failrec->len, pg_offset);
2367
2368         return bio;
2369 }
2370
2371 /*
2372  * this is a generic handler for readpage errors (default
2373  * readpage_io_failed_hook). if other copies exist, read those and write back
2374  * good data to the failed position. does not investigate in remapping the
2375  * failed extent elsewhere, hoping the device will be smart enough to do this as
2376  * needed
2377  */
2378
2379 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2380                               struct page *page, u64 start, u64 end,
2381                               int failed_mirror)
2382 {
2383         struct io_failure_record *failrec;
2384         struct inode *inode = page->mapping->host;
2385         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2386         struct bio *bio;
2387         int read_mode = 0;
2388         int ret;
2389
2390         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2391
2392         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2393         if (ret)
2394                 return ret;
2395
2396         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2397         if (!ret) {
2398                 free_io_failure(BTRFS_I(inode), failrec);
2399                 return -EIO;
2400         }
2401
2402         if (failed_bio->bi_vcnt > 1)
2403                 read_mode |= REQ_FAILFAST_DEV;
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(BTRFS_I(inode), failrec);
2412                 return -EIO;
2413         }
2414         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2415
2416         btrfs_debug(btrfs_sb(inode->i_sb),
2417                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2418                 read_mode, failrec->this_mirror, failrec->in_validation);
2419
2420         ret = tree->ops->submit_bio_hook(inode, bio, failrec->this_mirror,
2421                                          failrec->bio_flags, 0);
2422         if (ret) {
2423                 free_io_failure(BTRFS_I(inode), failrec);
2424                 bio_put(bio);
2425         }
2426
2427         return ret;
2428 }
2429
2430 /* lots and lots of room for performance fixes in the end_bio funcs */
2431
2432 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2433 {
2434         int uptodate = (err == 0);
2435         struct extent_io_tree *tree;
2436         int ret = 0;
2437
2438         tree = &BTRFS_I(page->mapping->host)->io_tree;
2439
2440         if (tree->ops && tree->ops->writepage_end_io_hook)
2441                 tree->ops->writepage_end_io_hook(page, start, end, NULL,
2442                                 uptodate);
2443
2444         if (!uptodate) {
2445                 ClearPageUptodate(page);
2446                 SetPageError(page);
2447                 ret = ret < 0 ? ret : -EIO;
2448                 mapping_set_error(page->mapping, ret);
2449         }
2450 }
2451
2452 /*
2453  * after a writepage IO is done, we need to:
2454  * clear the uptodate bits on error
2455  * clear the writeback bits in the extent tree for this IO
2456  * end_page_writeback if the page has no more pending IO
2457  *
2458  * Scheduling is not allowed, so the extent state tree is expected
2459  * to have one and only one object corresponding to this IO.
2460  */
2461 static void end_bio_extent_writepage(struct bio *bio)
2462 {
2463         struct bio_vec *bvec;
2464         u64 start;
2465         u64 end;
2466         int i;
2467
2468         bio_for_each_segment_all(bvec, bio, i) {
2469                 struct page *page = bvec->bv_page;
2470                 struct inode *inode = page->mapping->host;
2471                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
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(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(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                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2542
2543                 btrfs_debug(fs_info,
2544                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2545                         (u64)bio->bi_iter.bi_sector, bio->bi_error,
2546                         io_bio->mirror_num);
2547                 tree = &BTRFS_I(inode)->io_tree;
2548
2549                 /* We always issue full-page reads, but if some block
2550                  * in a page fails to read, blk_update_request() will
2551                  * advance bv_offset and adjust bv_len to compensate.
2552                  * Print a warning for nonzero offsets, and an error
2553                  * if they don't add up to a full page.  */
2554                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2555                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2556                                 btrfs_err(fs_info,
2557                                         "partial page read in btrfs with offset %u and length %u",
2558                                         bvec->bv_offset, bvec->bv_len);
2559                         else
2560                                 btrfs_info(fs_info,
2561                                         "incomplete page read in btrfs with offset %u and length %u",
2562                                         bvec->bv_offset, bvec->bv_len);
2563                 }
2564
2565                 start = page_offset(page);
2566                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2567                 len = bvec->bv_len;
2568
2569                 mirror = io_bio->mirror_num;
2570                 if (likely(uptodate && tree->ops)) {
2571                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2572                                                               page, start, end,
2573                                                               mirror);
2574                         if (ret)
2575                                 uptodate = 0;
2576                         else
2577                                 clean_io_failure(BTRFS_I(inode), start,
2578                                                 page, 0);
2579                 }
2580
2581                 if (likely(uptodate))
2582                         goto readpage_ok;
2583
2584                 if (tree->ops) {
2585                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2586                         if (!ret && !bio->bi_error)
2587                                 uptodate = 1;
2588                 } else {
2589                         /*
2590                          * The generic bio_readpage_error handles errors the
2591                          * following way: If possible, new read requests are
2592                          * created and submitted and will end up in
2593                          * end_bio_extent_readpage as well (if we're lucky, not
2594                          * in the !uptodate case). In that case it returns 0 and
2595                          * we just go on with the next page in our bio. If it
2596                          * can't handle the error it will return -EIO and we
2597                          * remain responsible for that page.
2598                          */
2599                         ret = bio_readpage_error(bio, offset, page, start, end,
2600                                                  mirror);
2601                         if (ret == 0) {
2602                                 uptodate = !bio->bi_error;
2603                                 offset += len;
2604                                 continue;
2605                         }
2606                 }
2607 readpage_ok:
2608                 if (likely(uptodate)) {
2609                         loff_t i_size = i_size_read(inode);
2610                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2611                         unsigned off;
2612
2613                         /* Zero out the end if this page straddles i_size */
2614                         off = i_size & (PAGE_SIZE-1);
2615                         if (page->index == end_index && off)
2616                                 zero_user_segment(page, off, PAGE_SIZE);
2617                         SetPageUptodate(page);
2618                 } else {
2619                         ClearPageUptodate(page);
2620                         SetPageError(page);
2621                 }
2622                 unlock_page(page);
2623                 offset += len;
2624
2625                 if (unlikely(!uptodate)) {
2626                         if (extent_len) {
2627                                 endio_readpage_release_extent(tree,
2628                                                               extent_start,
2629                                                               extent_len, 1);
2630                                 extent_start = 0;
2631                                 extent_len = 0;
2632                         }
2633                         endio_readpage_release_extent(tree, start,
2634                                                       end - start + 1, 0);
2635                 } else if (!extent_len) {
2636                         extent_start = start;
2637                         extent_len = end + 1 - start;
2638                 } else if (extent_start + extent_len == start) {
2639                         extent_len += end + 1 - start;
2640                 } else {
2641                         endio_readpage_release_extent(tree, extent_start,
2642                                                       extent_len, uptodate);
2643                         extent_start = start;
2644                         extent_len = end + 1 - start;
2645                 }
2646         }
2647
2648         if (extent_len)
2649                 endio_readpage_release_extent(tree, extent_start, extent_len,
2650                                               uptodate);
2651         if (io_bio->end_io)
2652                 io_bio->end_io(io_bio, bio->bi_error);
2653         bio_put(bio);
2654 }
2655
2656 /*
2657  * this allocates from the btrfs_bioset.  We're returning a bio right now
2658  * but you can call btrfs_io_bio for the appropriate container_of magic
2659  */
2660 struct bio *
2661 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2662                 gfp_t gfp_flags)
2663 {
2664         struct btrfs_io_bio *btrfs_bio;
2665         struct bio *bio;
2666
2667         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2668
2669         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2670                 while (!bio && (nr_vecs /= 2)) {
2671                         bio = bio_alloc_bioset(gfp_flags,
2672                                                nr_vecs, btrfs_bioset);
2673                 }
2674         }
2675
2676         if (bio) {
2677                 bio->bi_bdev = bdev;
2678                 bio->bi_iter.bi_sector = first_sector;
2679                 btrfs_bio = btrfs_io_bio(bio);
2680                 btrfs_bio->csum = NULL;
2681                 btrfs_bio->csum_allocated = NULL;
2682                 btrfs_bio->end_io = NULL;
2683         }
2684         return bio;
2685 }
2686
2687 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2688 {
2689         struct btrfs_io_bio *btrfs_bio;
2690         struct bio *new;
2691
2692         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2693         if (new) {
2694                 btrfs_bio = btrfs_io_bio(new);
2695                 btrfs_bio->csum = NULL;
2696                 btrfs_bio->csum_allocated = NULL;
2697                 btrfs_bio->end_io = NULL;
2698         }
2699         return new;
2700 }
2701
2702 /* this also allocates from the btrfs_bioset */
2703 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2704 {
2705         struct btrfs_io_bio *btrfs_bio;
2706         struct bio *bio;
2707
2708         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2709         if (bio) {
2710                 btrfs_bio = btrfs_io_bio(bio);
2711                 btrfs_bio->csum = NULL;
2712                 btrfs_bio->csum_allocated = NULL;
2713                 btrfs_bio->end_io = NULL;
2714         }
2715         return bio;
2716 }
2717
2718
2719 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2720                                        unsigned long bio_flags)
2721 {
2722         int ret = 0;
2723         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2724         struct page *page = bvec->bv_page;
2725         struct extent_io_tree *tree = bio->bi_private;
2726         u64 start;
2727
2728         start = page_offset(page) + bvec->bv_offset;
2729
2730         bio->bi_private = NULL;
2731         bio_get(bio);
2732
2733         if (tree->ops)
2734                 ret = tree->ops->submit_bio_hook(page->mapping->host, bio,
2735                                            mirror_num, bio_flags, start);
2736         else
2737                 btrfsic_submit_bio(bio);
2738
2739         bio_put(bio);
2740         return ret;
2741 }
2742
2743 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2744                      unsigned long offset, size_t size, struct bio *bio,
2745                      unsigned long bio_flags)
2746 {
2747         int ret = 0;
2748         if (tree->ops)
2749                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2750                                                 bio_flags);
2751         return ret;
2752
2753 }
2754
2755 static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree,
2756                               struct writeback_control *wbc,
2757                               struct page *page, sector_t sector,
2758                               size_t size, unsigned long offset,
2759                               struct block_device *bdev,
2760                               struct bio **bio_ret,
2761                               bio_end_io_t end_io_func,
2762                               int mirror_num,
2763                               unsigned long prev_bio_flags,
2764                               unsigned long bio_flags,
2765                               bool force_bio_submit)
2766 {
2767         int ret = 0;
2768         struct bio *bio;
2769         int contig = 0;
2770         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2771         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2772
2773         if (bio_ret && *bio_ret) {
2774                 bio = *bio_ret;
2775                 if (old_compressed)
2776                         contig = bio->bi_iter.bi_sector == sector;
2777                 else
2778                         contig = bio_end_sector(bio) == sector;
2779
2780                 if (prev_bio_flags != bio_flags || !contig ||
2781                     force_bio_submit ||
2782                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2783                     bio_add_page(bio, page, page_size, offset) < page_size) {
2784                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2785                         if (ret < 0) {
2786                                 *bio_ret = NULL;
2787                                 return ret;
2788                         }
2789                         bio = NULL;
2790                 } else {
2791                         if (wbc)
2792                                 wbc_account_io(wbc, page, page_size);
2793                         return 0;
2794                 }
2795         }
2796
2797         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2798                         GFP_NOFS | __GFP_HIGH);
2799         if (!bio)
2800                 return -ENOMEM;
2801
2802         bio_add_page(bio, page, page_size, offset);
2803         bio->bi_end_io = end_io_func;
2804         bio->bi_private = tree;
2805         bio_set_op_attrs(bio, op, op_flags);
2806         if (wbc) {
2807                 wbc_init_bio(wbc, bio);
2808                 wbc_account_io(wbc, page, page_size);
2809         }
2810
2811         if (bio_ret)
2812                 *bio_ret = bio;
2813         else
2814                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2815
2816         return ret;
2817 }
2818
2819 static void attach_extent_buffer_page(struct extent_buffer *eb,
2820                                       struct page *page)
2821 {
2822         if (!PagePrivate(page)) {
2823                 SetPagePrivate(page);
2824                 get_page(page);
2825                 set_page_private(page, (unsigned long)eb);
2826         } else {
2827                 WARN_ON(page->private != (unsigned long)eb);
2828         }
2829 }
2830
2831 void set_page_extent_mapped(struct page *page)
2832 {
2833         if (!PagePrivate(page)) {
2834                 SetPagePrivate(page);
2835                 get_page(page);
2836                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2837         }
2838 }
2839
2840 static struct extent_map *
2841 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2842                  u64 start, u64 len, get_extent_t *get_extent,
2843                  struct extent_map **em_cached)
2844 {
2845         struct extent_map *em;
2846
2847         if (em_cached && *em_cached) {
2848                 em = *em_cached;
2849                 if (extent_map_in_tree(em) && start >= em->start &&
2850                     start < extent_map_end(em)) {
2851                         atomic_inc(&em->refs);
2852                         return em;
2853                 }
2854
2855                 free_extent_map(em);
2856                 *em_cached = NULL;
2857         }
2858
2859         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2860         if (em_cached && !IS_ERR_OR_NULL(em)) {
2861                 BUG_ON(*em_cached);
2862                 atomic_inc(&em->refs);
2863                 *em_cached = em;
2864         }
2865         return em;
2866 }
2867 /*
2868  * basic readpage implementation.  Locked extent state structs are inserted
2869  * into the tree that are removed when the IO is done (by the end_io
2870  * handlers)
2871  * XXX JDM: This needs looking at to ensure proper page locking
2872  * return 0 on success, otherwise return error
2873  */
2874 static int __do_readpage(struct extent_io_tree *tree,
2875                          struct page *page,
2876                          get_extent_t *get_extent,
2877                          struct extent_map **em_cached,
2878                          struct bio **bio, int mirror_num,
2879                          unsigned long *bio_flags, int read_flags,
2880                          u64 *prev_em_start)
2881 {
2882         struct inode *inode = page->mapping->host;
2883         u64 start = page_offset(page);
2884         u64 page_end = start + PAGE_SIZE - 1;
2885         u64 end;
2886         u64 cur = start;
2887         u64 extent_offset;
2888         u64 last_byte = i_size_read(inode);
2889         u64 block_start;
2890         u64 cur_end;
2891         sector_t sector;
2892         struct extent_map *em;
2893         struct block_device *bdev;
2894         int ret = 0;
2895         int nr = 0;
2896         size_t pg_offset = 0;
2897         size_t iosize;
2898         size_t disk_io_size;
2899         size_t blocksize = inode->i_sb->s_blocksize;
2900         unsigned long this_bio_flag = 0;
2901
2902         set_page_extent_mapped(page);
2903
2904         end = page_end;
2905         if (!PageUptodate(page)) {
2906                 if (cleancache_get_page(page) == 0) {
2907                         BUG_ON(blocksize != PAGE_SIZE);
2908                         unlock_extent(tree, start, end);
2909                         goto out;
2910                 }
2911         }
2912
2913         if (page->index == last_byte >> PAGE_SHIFT) {
2914                 char *userpage;
2915                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2916
2917                 if (zero_offset) {
2918                         iosize = PAGE_SIZE - zero_offset;
2919                         userpage = kmap_atomic(page);
2920                         memset(userpage + zero_offset, 0, iosize);
2921                         flush_dcache_page(page);
2922                         kunmap_atomic(userpage);
2923                 }
2924         }
2925         while (cur <= end) {
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                 ret = submit_extent_page(REQ_OP_READ, read_flags, tree, NULL,
3061                                          page, sector, disk_io_size, pg_offset,
3062                                          bdev, bio,
3063                                          end_bio_extent_readpage, mirror_num,
3064                                          *bio_flags,
3065                                          this_bio_flag,
3066                                          force_bio_submit);
3067                 if (!ret) {
3068                         nr++;
3069                         *bio_flags = this_bio_flag;
3070                 } else {
3071                         SetPageError(page);
3072                         unlock_extent(tree, cur, cur + iosize - 1);
3073                         goto out;
3074                 }
3075                 cur = cur + iosize;
3076                 pg_offset += iosize;
3077         }
3078 out:
3079         if (!nr) {
3080                 if (!PageError(page))
3081                         SetPageUptodate(page);
3082                 unlock_page(page);
3083         }
3084         return ret;
3085 }
3086
3087 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3088                                              struct page *pages[], int nr_pages,
3089                                              u64 start, u64 end,
3090                                              get_extent_t *get_extent,
3091                                              struct extent_map **em_cached,
3092                                              struct bio **bio, int mirror_num,
3093                                              unsigned long *bio_flags,
3094                                              u64 *prev_em_start)
3095 {
3096         struct inode *inode;
3097         struct btrfs_ordered_extent *ordered;
3098         int index;
3099
3100         inode = pages[0]->mapping->host;
3101         while (1) {
3102                 lock_extent(tree, start, end);
3103                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3104                                                      end - start + 1);
3105                 if (!ordered)
3106                         break;
3107                 unlock_extent(tree, start, end);
3108                 btrfs_start_ordered_extent(inode, ordered, 1);
3109                 btrfs_put_ordered_extent(ordered);
3110         }
3111
3112         for (index = 0; index < nr_pages; index++) {
3113                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3114                               mirror_num, bio_flags, 0, prev_em_start);
3115                 put_page(pages[index]);
3116         }
3117 }
3118
3119 static void __extent_readpages(struct extent_io_tree *tree,
3120                                struct page *pages[],
3121                                int nr_pages, get_extent_t *get_extent,
3122                                struct extent_map **em_cached,
3123                                struct bio **bio, int mirror_num,
3124                                unsigned long *bio_flags,
3125                                u64 *prev_em_start)
3126 {
3127         u64 start = 0;
3128         u64 end = 0;
3129         u64 page_start;
3130         int index;
3131         int first_index = 0;
3132
3133         for (index = 0; index < nr_pages; index++) {
3134                 page_start = page_offset(pages[index]);
3135                 if (!end) {
3136                         start = page_start;
3137                         end = start + PAGE_SIZE - 1;
3138                         first_index = index;
3139                 } else if (end + 1 == page_start) {
3140                         end += PAGE_SIZE;
3141                 } else {
3142                         __do_contiguous_readpages(tree, &pages[first_index],
3143                                                   index - first_index, start,
3144                                                   end, get_extent, em_cached,
3145                                                   bio, mirror_num, bio_flags,
3146                                                   prev_em_start);
3147                         start = page_start;
3148                         end = start + PAGE_SIZE - 1;
3149                         first_index = index;
3150                 }
3151         }
3152
3153         if (end)
3154                 __do_contiguous_readpages(tree, &pages[first_index],
3155                                           index - first_index, start,
3156                                           end, get_extent, em_cached, bio,
3157                                           mirror_num, bio_flags,
3158                                           prev_em_start);
3159 }
3160
3161 static int __extent_read_full_page(struct extent_io_tree *tree,
3162                                    struct page *page,
3163                                    get_extent_t *get_extent,
3164                                    struct bio **bio, int mirror_num,
3165                                    unsigned long *bio_flags, int read_flags)
3166 {
3167         struct inode *inode = page->mapping->host;
3168         struct btrfs_ordered_extent *ordered;
3169         u64 start = page_offset(page);
3170         u64 end = start + PAGE_SIZE - 1;
3171         int ret;
3172
3173         while (1) {
3174                 lock_extent(tree, start, end);
3175                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3176                                                 PAGE_SIZE);
3177                 if (!ordered)
3178                         break;
3179                 unlock_extent(tree, start, end);
3180                 btrfs_start_ordered_extent(inode, ordered, 1);
3181                 btrfs_put_ordered_extent(ordered);
3182         }
3183
3184         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3185                             bio_flags, read_flags, NULL);
3186         return ret;
3187 }
3188
3189 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3190                             get_extent_t *get_extent, int mirror_num)
3191 {
3192         struct bio *bio = NULL;
3193         unsigned long bio_flags = 0;
3194         int ret;
3195
3196         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3197                                       &bio_flags, 0);
3198         if (bio)
3199                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3200         return ret;
3201 }
3202
3203 static void update_nr_written(struct writeback_control *wbc,
3204                               unsigned long nr_written)
3205 {
3206         wbc->nr_to_write -= nr_written;
3207 }
3208
3209 /*
3210  * helper for __extent_writepage, doing all of the delayed allocation setup.
3211  *
3212  * This returns 1 if our fill_delalloc function did all the work required
3213  * to write the page (copy into inline extent).  In this case the IO has
3214  * been started and the page is already unlocked.
3215  *
3216  * This returns 0 if all went well (page still locked)
3217  * This returns < 0 if there were errors (page still locked)
3218  */
3219 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3220                               struct page *page, struct writeback_control *wbc,
3221                               struct extent_page_data *epd,
3222                               u64 delalloc_start,
3223                               unsigned long *nr_written)
3224 {
3225         struct extent_io_tree *tree = epd->tree;
3226         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3227         u64 nr_delalloc;
3228         u64 delalloc_to_write = 0;
3229         u64 delalloc_end = 0;
3230         int ret;
3231         int page_started = 0;
3232
3233         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3234                 return 0;
3235
3236         while (delalloc_end < page_end) {
3237                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3238                                                page,
3239                                                &delalloc_start,
3240                                                &delalloc_end,
3241                                                BTRFS_MAX_EXTENT_SIZE);
3242                 if (nr_delalloc == 0) {
3243                         delalloc_start = delalloc_end + 1;
3244                         continue;
3245                 }
3246                 ret = tree->ops->fill_delalloc(inode, page,
3247                                                delalloc_start,
3248                                                delalloc_end,
3249                                                &page_started,
3250                                                nr_written);
3251                 /* File system has been set read-only */
3252                 if (ret) {
3253                         SetPageError(page);
3254                         /* fill_delalloc should be return < 0 for error
3255                          * but just in case, we use > 0 here meaning the
3256                          * IO is started, so we don't want to return > 0
3257                          * unless things are going well.
3258                          */
3259                         ret = ret < 0 ? ret : -EIO;
3260                         goto done;
3261                 }
3262                 /*
3263                  * delalloc_end is already one less than the total length, so
3264                  * we don't subtract one from PAGE_SIZE
3265                  */
3266                 delalloc_to_write += (delalloc_end - delalloc_start +
3267                                       PAGE_SIZE) >> PAGE_SHIFT;
3268                 delalloc_start = delalloc_end + 1;
3269         }
3270         if (wbc->nr_to_write < delalloc_to_write) {
3271                 int thresh = 8192;
3272
3273                 if (delalloc_to_write < thresh * 2)
3274                         thresh = delalloc_to_write;
3275                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3276                                          thresh);
3277         }
3278
3279         /* did the fill delalloc function already unlock and start
3280          * the IO?
3281          */
3282         if (page_started) {
3283                 /*
3284                  * we've unlocked the page, so we can't update
3285                  * the mapping's writeback index, just update
3286                  * nr_to_write.
3287                  */
3288                 wbc->nr_to_write -= *nr_written;
3289                 return 1;
3290         }
3291
3292         ret = 0;
3293
3294 done:
3295         return ret;
3296 }
3297
3298 /*
3299  * helper for __extent_writepage.  This calls the writepage start hooks,
3300  * and does the loop to map the page into extents and bios.
3301  *
3302  * We return 1 if the IO is started and the page is unlocked,
3303  * 0 if all went well (page still locked)
3304  * < 0 if there were errors (page still locked)
3305  */
3306 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3307                                  struct page *page,
3308                                  struct writeback_control *wbc,
3309                                  struct extent_page_data *epd,
3310                                  loff_t i_size,
3311                                  unsigned long nr_written,
3312                                  int write_flags, int *nr_ret)
3313 {
3314         struct extent_io_tree *tree = epd->tree;
3315         u64 start = page_offset(page);
3316         u64 page_end = start + PAGE_SIZE - 1;
3317         u64 end;
3318         u64 cur = start;
3319         u64 extent_offset;
3320         u64 block_start;
3321         u64 iosize;
3322         sector_t sector;
3323         struct extent_map *em;
3324         struct block_device *bdev;
3325         size_t pg_offset = 0;
3326         size_t blocksize;
3327         int ret = 0;
3328         int nr = 0;
3329         bool compressed;
3330
3331         if (tree->ops && tree->ops->writepage_start_hook) {
3332                 ret = tree->ops->writepage_start_hook(page, start,
3333                                                       page_end);
3334                 if (ret) {
3335                         /* Fixup worker will requeue */
3336                         if (ret == -EBUSY)
3337                                 wbc->pages_skipped++;
3338                         else
3339                                 redirty_page_for_writepage(wbc, page);
3340
3341                         update_nr_written(wbc, nr_written);
3342                         unlock_page(page);
3343                         return 1;
3344                 }
3345         }
3346
3347         /*
3348          * we don't want to touch the inode after unlocking the page,
3349          * so we update the mapping writeback index now
3350          */
3351         update_nr_written(wbc, nr_written + 1);
3352
3353         end = page_end;
3354         if (i_size <= start) {
3355                 if (tree->ops && tree->ops->writepage_end_io_hook)
3356                         tree->ops->writepage_end_io_hook(page, start,
3357                                                          page_end, NULL, 1);
3358                 goto done;
3359         }
3360
3361         blocksize = inode->i_sb->s_blocksize;
3362
3363         while (cur <= end) {
3364                 u64 em_end;
3365
3366                 if (cur >= i_size) {
3367                         if (tree->ops && tree->ops->writepage_end_io_hook)
3368                                 tree->ops->writepage_end_io_hook(page, cur,
3369                                                          page_end, NULL, 1);
3370                         break;
3371                 }
3372                 em = epd->get_extent(BTRFS_I(inode), page, pg_offset, cur,
3373                                      end - cur + 1, 1);
3374                 if (IS_ERR_OR_NULL(em)) {
3375                         SetPageError(page);
3376                         ret = PTR_ERR_OR_ZERO(em);
3377                         break;
3378                 }
3379
3380                 extent_offset = cur - em->start;
3381                 em_end = extent_map_end(em);
3382                 BUG_ON(em_end <= cur);
3383                 BUG_ON(end < cur);
3384                 iosize = min(em_end - cur, end - cur + 1);
3385                 iosize = ALIGN(iosize, blocksize);
3386                 sector = (em->block_start + extent_offset) >> 9;
3387                 bdev = em->bdev;
3388                 block_start = em->block_start;
3389                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3390                 free_extent_map(em);
3391                 em = NULL;
3392
3393                 /*
3394                  * compressed and inline extents are written through other
3395                  * paths in the FS
3396                  */
3397                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3398                     block_start == EXTENT_MAP_INLINE) {
3399                         /*
3400                          * end_io notification does not happen here for
3401                          * compressed extents
3402                          */
3403                         if (!compressed && tree->ops &&
3404                             tree->ops->writepage_end_io_hook)
3405                                 tree->ops->writepage_end_io_hook(page, cur,
3406                                                          cur + iosize - 1,
3407                                                          NULL, 1);
3408                         else if (compressed) {
3409                                 /* we don't want to end_page_writeback on
3410                                  * a compressed extent.  this happens
3411                                  * elsewhere
3412                                  */
3413                                 nr++;
3414                         }
3415
3416                         cur += iosize;
3417                         pg_offset += iosize;
3418                         continue;
3419                 }
3420
3421                 set_range_writeback(tree, cur, cur + iosize - 1);
3422                 if (!PageWriteback(page)) {
3423                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3424                                    "page %lu not writeback, cur %llu end %llu",
3425                                page->index, cur, end);
3426                 }
3427
3428                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3429                                          page, sector, iosize, pg_offset,
3430                                          bdev, &epd->bio,
3431                                          end_bio_extent_writepage,
3432                                          0, 0, 0, false);
3433                 if (ret) {
3434                         SetPageError(page);
3435                         if (PageWriteback(page))
3436                                 end_page_writeback(page);
3437                 }
3438
3439                 cur = cur + iosize;
3440                 pg_offset += iosize;
3441                 nr++;
3442         }
3443 done:
3444         *nr_ret = nr;
3445         return ret;
3446 }
3447
3448 /*
3449  * the writepage semantics are similar to regular writepage.  extent
3450  * records are inserted to lock ranges in the tree, and as dirty areas
3451  * are found, they are marked writeback.  Then the lock bits are removed
3452  * and the end_io handler clears the writeback ranges
3453  */
3454 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3455                               void *data)
3456 {
3457         struct inode *inode = page->mapping->host;
3458         struct extent_page_data *epd = data;
3459         u64 start = page_offset(page);
3460         u64 page_end = start + PAGE_SIZE - 1;
3461         int ret;
3462         int nr = 0;
3463         size_t pg_offset = 0;
3464         loff_t i_size = i_size_read(inode);
3465         unsigned long end_index = i_size >> PAGE_SHIFT;
3466         int write_flags = 0;
3467         unsigned long nr_written = 0;
3468
3469         if (wbc->sync_mode == WB_SYNC_ALL)
3470                 write_flags = REQ_SYNC;
3471
3472         trace___extent_writepage(page, inode, wbc);
3473
3474         WARN_ON(!PageLocked(page));
3475
3476         ClearPageError(page);
3477
3478         pg_offset = i_size & (PAGE_SIZE - 1);
3479         if (page->index > end_index ||
3480            (page->index == end_index && !pg_offset)) {
3481                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3482                 unlock_page(page);
3483                 return 0;
3484         }
3485
3486         if (page->index == end_index) {
3487                 char *userpage;
3488
3489                 userpage = kmap_atomic(page);
3490                 memset(userpage + pg_offset, 0,
3491                        PAGE_SIZE - pg_offset);
3492                 kunmap_atomic(userpage);
3493                 flush_dcache_page(page);
3494         }
3495
3496         pg_offset = 0;
3497
3498         set_page_extent_mapped(page);
3499
3500         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3501         if (ret == 1)
3502                 goto done_unlocked;
3503         if (ret)
3504                 goto done;
3505
3506         ret = __extent_writepage_io(inode, page, wbc, epd,
3507                                     i_size, nr_written, write_flags, &nr);
3508         if (ret == 1)
3509                 goto done_unlocked;
3510
3511 done:
3512         if (nr == 0) {
3513                 /* make sure the mapping tag for page dirty gets cleared */
3514                 set_page_writeback(page);
3515                 end_page_writeback(page);
3516         }
3517         if (PageError(page)) {
3518                 ret = ret < 0 ? ret : -EIO;
3519                 end_extent_writepage(page, ret, start, page_end);
3520         }
3521         unlock_page(page);
3522         return ret;
3523
3524 done_unlocked:
3525         return 0;
3526 }
3527
3528 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3529 {
3530         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3531                        TASK_UNINTERRUPTIBLE);
3532 }
3533
3534 static noinline_for_stack int
3535 lock_extent_buffer_for_io(struct extent_buffer *eb,
3536                           struct btrfs_fs_info *fs_info,
3537                           struct extent_page_data *epd)
3538 {
3539         unsigned long i, num_pages;
3540         int flush = 0;
3541         int ret = 0;
3542
3543         if (!btrfs_try_tree_write_lock(eb)) {
3544                 flush = 1;
3545                 flush_write_bio(epd);
3546                 btrfs_tree_lock(eb);
3547         }
3548
3549         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3550                 btrfs_tree_unlock(eb);
3551                 if (!epd->sync_io)
3552                         return 0;
3553                 if (!flush) {
3554                         flush_write_bio(epd);
3555                         flush = 1;
3556                 }
3557                 while (1) {
3558                         wait_on_extent_buffer_writeback(eb);
3559                         btrfs_tree_lock(eb);
3560                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3561                                 break;
3562                         btrfs_tree_unlock(eb);
3563                 }
3564         }
3565
3566         /*
3567          * We need to do this to prevent races in people who check if the eb is
3568          * under IO since we can end up having no IO bits set for a short period
3569          * of time.
3570          */
3571         spin_lock(&eb->refs_lock);
3572         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3573                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3574                 spin_unlock(&eb->refs_lock);
3575                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3576                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3577                                      -eb->len,
3578                                      fs_info->dirty_metadata_batch);
3579                 ret = 1;
3580         } else {
3581                 spin_unlock(&eb->refs_lock);
3582         }
3583
3584         btrfs_tree_unlock(eb);
3585
3586         if (!ret)
3587                 return ret;
3588
3589         num_pages = num_extent_pages(eb->start, eb->len);
3590         for (i = 0; i < num_pages; i++) {
3591                 struct page *p = eb->pages[i];
3592
3593                 if (!trylock_page(p)) {
3594                         if (!flush) {
3595                                 flush_write_bio(epd);
3596                                 flush = 1;
3597                         }
3598                         lock_page(p);
3599                 }
3600         }
3601
3602         return ret;
3603 }
3604
3605 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3606 {
3607         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3608         smp_mb__after_atomic();
3609         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3610 }
3611
3612 static void set_btree_ioerr(struct page *page)
3613 {
3614         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3615
3616         SetPageError(page);
3617         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3618                 return;
3619
3620         /*
3621          * If writeback for a btree extent that doesn't belong to a log tree
3622          * failed, increment the counter transaction->eb_write_errors.
3623          * We do this because while the transaction is running and before it's
3624          * committing (when we call filemap_fdata[write|wait]_range against
3625          * the btree inode), we might have
3626          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3627          * returns an error or an error happens during writeback, when we're
3628          * committing the transaction we wouldn't know about it, since the pages
3629          * can be no longer dirty nor marked anymore for writeback (if a
3630          * subsequent modification to the extent buffer didn't happen before the
3631          * transaction commit), which makes filemap_fdata[write|wait]_range not
3632          * able to find the pages tagged with SetPageError at transaction
3633          * commit time. So if this happens we must abort the transaction,
3634          * otherwise we commit a super block with btree roots that point to
3635          * btree nodes/leafs whose content on disk is invalid - either garbage
3636          * or the content of some node/leaf from a past generation that got
3637          * cowed or deleted and is no longer valid.
3638          *
3639          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3640          * not be enough - we need to distinguish between log tree extents vs
3641          * non-log tree extents, and the next filemap_fdatawait_range() call
3642          * will catch and clear such errors in the mapping - and that call might
3643          * be from a log sync and not from a transaction commit. Also, checking
3644          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3645          * not done and would not be reliable - the eb might have been released
3646          * from memory and reading it back again means that flag would not be
3647          * set (since it's a runtime flag, not persisted on disk).
3648          *
3649          * Using the flags below in the btree inode also makes us achieve the
3650          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3651          * writeback for all dirty pages and before filemap_fdatawait_range()
3652          * is called, the writeback for all dirty pages had already finished
3653          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3654          * filemap_fdatawait_range() would return success, as it could not know
3655          * that writeback errors happened (the pages were no longer tagged for
3656          * writeback).
3657          */
3658         switch (eb->log_index) {
3659         case -1:
3660                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3661                 break;
3662         case 0:
3663                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3664                 break;
3665         case 1:
3666                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3667                 break;
3668         default:
3669                 BUG(); /* unexpected, logic error */
3670         }
3671 }
3672
3673 static void end_bio_extent_buffer_writepage(struct bio *bio)
3674 {
3675         struct bio_vec *bvec;
3676         struct extent_buffer *eb;
3677         int i, done;
3678
3679         bio_for_each_segment_all(bvec, bio, i) {
3680                 struct page *page = bvec->bv_page;
3681
3682                 eb = (struct extent_buffer *)page->private;
3683                 BUG_ON(!eb);
3684                 done = atomic_dec_and_test(&eb->io_pages);
3685
3686                 if (bio->bi_error ||
3687                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3688                         ClearPageUptodate(page);
3689                         set_btree_ioerr(page);
3690                 }
3691
3692                 end_page_writeback(page);
3693
3694                 if (!done)
3695                         continue;
3696
3697                 end_extent_buffer_writeback(eb);
3698         }
3699
3700         bio_put(bio);
3701 }
3702
3703 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3704                         struct btrfs_fs_info *fs_info,
3705                         struct writeback_control *wbc,
3706                         struct extent_page_data *epd)
3707 {
3708         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3709         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3710         u64 offset = eb->start;
3711         u32 nritems;
3712         unsigned long i, num_pages;
3713         unsigned long bio_flags = 0;
3714         unsigned long start, end;
3715         int write_flags = (epd->sync_io ? REQ_SYNC : 0) | REQ_META;
3716         int ret = 0;
3717
3718         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3719         num_pages = num_extent_pages(eb->start, eb->len);
3720         atomic_set(&eb->io_pages, num_pages);
3721         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3722                 bio_flags = EXTENT_BIO_TREE_LOG;
3723
3724         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3725         nritems = btrfs_header_nritems(eb);
3726         if (btrfs_header_level(eb) > 0) {
3727                 end = btrfs_node_key_ptr_offset(nritems);
3728
3729                 memzero_extent_buffer(eb, end, eb->len - end);
3730         } else {
3731                 /*
3732                  * leaf:
3733                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3734                  */
3735                 start = btrfs_item_nr_offset(nritems);
3736                 end = btrfs_leaf_data(eb) + leaf_data_end(fs_info, eb);
3737                 memzero_extent_buffer(eb, start, end - start);
3738         }
3739
3740         for (i = 0; i < num_pages; i++) {
3741                 struct page *p = eb->pages[i];
3742
3743                 clear_page_dirty_for_io(p);
3744                 set_page_writeback(p);
3745                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3746                                          p, offset >> 9, PAGE_SIZE, 0, bdev,
3747                                          &epd->bio,
3748                                          end_bio_extent_buffer_writepage,
3749                                          0, epd->bio_flags, bio_flags, false);
3750                 epd->bio_flags = bio_flags;
3751                 if (ret) {
3752                         set_btree_ioerr(p);
3753                         if (PageWriteback(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(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 address_space *mapping,
3914                              struct writeback_control *wbc,
3915                              writepage_t writepage, void *data,
3916                              void (*flush_fn)(void *))
3917 {
3918         struct inode *inode = mapping->host;
3919         int ret = 0;
3920         int done = 0;
3921         int nr_to_write_done = 0;
3922         struct pagevec pvec;
3923         int nr_pages;
3924         pgoff_t index;
3925         pgoff_t end;            /* Inclusive */
3926         pgoff_t done_index;
3927         int range_whole = 0;
3928         int scanned = 0;
3929         int tag;
3930
3931         /*
3932          * We have to hold onto the inode so that ordered extents can do their
3933          * work when the IO finishes.  The alternative to this is failing to add
3934          * an ordered extent if the igrab() fails there and that is a huge pain
3935          * to deal with, so instead just hold onto the inode throughout the
3936          * writepages operation.  If it fails here we are freeing up the inode
3937          * anyway and we'd rather not waste our time writing out stuff that is
3938          * going to be truncated anyway.
3939          */
3940         if (!igrab(inode))
3941                 return 0;
3942
3943         pagevec_init(&pvec, 0);
3944         if (wbc->range_cyclic) {
3945                 index = mapping->writeback_index; /* Start from prev offset */
3946                 end = -1;
3947         } else {
3948                 index = wbc->range_start >> PAGE_SHIFT;
3949                 end = wbc->range_end >> PAGE_SHIFT;
3950                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3951                         range_whole = 1;
3952                 scanned = 1;
3953         }
3954         if (wbc->sync_mode == WB_SYNC_ALL)
3955                 tag = PAGECACHE_TAG_TOWRITE;
3956         else
3957                 tag = PAGECACHE_TAG_DIRTY;
3958 retry:
3959         if (wbc->sync_mode == WB_SYNC_ALL)
3960                 tag_pages_for_writeback(mapping, index, end);
3961         done_index = index;
3962         while (!done && !nr_to_write_done && (index <= end) &&
3963                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3964                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3965                 unsigned i;
3966
3967                 scanned = 1;
3968                 for (i = 0; i < nr_pages; i++) {
3969                         struct page *page = pvec.pages[i];
3970
3971                         done_index = page->index;
3972                         /*
3973                          * At this point we hold neither mapping->tree_lock nor
3974                          * lock on the page itself: the page may be truncated or
3975                          * invalidated (changing page->mapping to NULL), or even
3976                          * swizzled back from swapper_space to tmpfs file
3977                          * mapping
3978                          */
3979                         if (!trylock_page(page)) {
3980                                 flush_fn(data);
3981                                 lock_page(page);
3982                         }
3983
3984                         if (unlikely(page->mapping != mapping)) {
3985                                 unlock_page(page);
3986                                 continue;
3987                         }
3988
3989                         if (!wbc->range_cyclic && page->index > end) {
3990                                 done = 1;
3991                                 unlock_page(page);
3992                                 continue;
3993                         }
3994
3995                         if (wbc->sync_mode != WB_SYNC_NONE) {
3996                                 if (PageWriteback(page))
3997                                         flush_fn(data);
3998                                 wait_on_page_writeback(page);
3999                         }
4000
4001                         if (PageWriteback(page) ||
4002                             !clear_page_dirty_for_io(page)) {
4003                                 unlock_page(page);
4004                                 continue;
4005                         }
4006
4007                         ret = (*writepage)(page, wbc, data);
4008
4009                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4010                                 unlock_page(page);
4011                                 ret = 0;
4012                         }
4013                         if (ret < 0) {
4014                                 /*
4015                                  * done_index is set past this page,
4016                                  * so media errors will not choke
4017                                  * background writeout for the entire
4018                                  * file. This has consequences for
4019                                  * range_cyclic semantics (ie. it may
4020                                  * not be suitable for data integrity
4021                                  * writeout).
4022                                  */
4023                                 done_index = page->index + 1;
4024                                 done = 1;
4025                                 break;
4026                         }
4027
4028                         /*
4029                          * the filesystem may choose to bump up nr_to_write.
4030                          * We have to make sure to honor the new nr_to_write
4031                          * at any time
4032                          */
4033                         nr_to_write_done = wbc->nr_to_write <= 0;
4034                 }
4035                 pagevec_release(&pvec);
4036                 cond_resched();
4037         }
4038         if (!scanned && !done) {
4039                 /*
4040                  * We hit the last page and there is more work to be done: wrap
4041                  * back to the start of the file
4042                  */
4043                 scanned = 1;
4044                 index = 0;
4045                 goto retry;
4046         }
4047
4048         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4049                 mapping->writeback_index = done_index;
4050
4051         btrfs_add_delayed_iput(inode);
4052         return ret;
4053 }
4054
4055 static void flush_epd_write_bio(struct extent_page_data *epd)
4056 {
4057         if (epd->bio) {
4058                 int ret;
4059
4060                 bio_set_op_attrs(epd->bio, REQ_OP_WRITE,
4061                                  epd->sync_io ? REQ_SYNC : 0);
4062
4063                 ret = submit_one_bio(epd->bio, 0, epd->bio_flags);
4064                 BUG_ON(ret < 0); /* -ENOMEM */
4065                 epd->bio = NULL;
4066         }
4067 }
4068
4069 static noinline void flush_write_bio(void *data)
4070 {
4071         struct extent_page_data *epd = data;
4072         flush_epd_write_bio(epd);
4073 }
4074
4075 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4076                           get_extent_t *get_extent,
4077                           struct writeback_control *wbc)
4078 {
4079         int ret;
4080         struct extent_page_data epd = {
4081                 .bio = NULL,
4082                 .tree = tree,
4083                 .get_extent = get_extent,
4084                 .extent_locked = 0,
4085                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4086                 .bio_flags = 0,
4087         };
4088
4089         ret = __extent_writepage(page, wbc, &epd);
4090
4091         flush_epd_write_bio(&epd);
4092         return ret;
4093 }
4094
4095 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4096                               u64 start, u64 end, get_extent_t *get_extent,
4097                               int mode)
4098 {
4099         int ret = 0;
4100         struct address_space *mapping = inode->i_mapping;
4101         struct page *page;
4102         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4103                 PAGE_SHIFT;
4104
4105         struct extent_page_data epd = {
4106                 .bio = NULL,
4107                 .tree = tree,
4108                 .get_extent = get_extent,
4109                 .extent_locked = 1,
4110                 .sync_io = mode == WB_SYNC_ALL,
4111                 .bio_flags = 0,
4112         };
4113         struct writeback_control wbc_writepages = {
4114                 .sync_mode      = mode,
4115                 .nr_to_write    = nr_pages * 2,
4116                 .range_start    = start,
4117                 .range_end      = end + 1,
4118         };
4119
4120         while (start <= end) {
4121                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4122                 if (clear_page_dirty_for_io(page))
4123                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4124                 else {
4125                         if (tree->ops && tree->ops->writepage_end_io_hook)
4126                                 tree->ops->writepage_end_io_hook(page, start,
4127                                                  start + PAGE_SIZE - 1,
4128                                                  NULL, 1);
4129                         unlock_page(page);
4130                 }
4131                 put_page(page);
4132                 start += PAGE_SIZE;
4133         }
4134
4135         flush_epd_write_bio(&epd);
4136         return ret;
4137 }
4138
4139 int extent_writepages(struct extent_io_tree *tree,
4140                       struct address_space *mapping,
4141                       get_extent_t *get_extent,
4142                       struct writeback_control *wbc)
4143 {
4144         int ret = 0;
4145         struct extent_page_data epd = {
4146                 .bio = NULL,
4147                 .tree = tree,
4148                 .get_extent = get_extent,
4149                 .extent_locked = 0,
4150                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4151                 .bio_flags = 0,
4152         };
4153
4154         ret = extent_write_cache_pages(mapping, wbc, __extent_writepage, &epd,
4155                                        flush_write_bio);
4156         flush_epd_write_bio(&epd);
4157         return ret;
4158 }
4159
4160 int extent_readpages(struct extent_io_tree *tree,
4161                      struct address_space *mapping,
4162                      struct list_head *pages, unsigned nr_pages,
4163                      get_extent_t get_extent)
4164 {
4165         struct bio *bio = NULL;
4166         unsigned page_idx;
4167         unsigned long bio_flags = 0;
4168         struct page *pagepool[16];
4169         struct page *page;
4170         struct extent_map *em_cached = NULL;
4171         int nr = 0;
4172         u64 prev_em_start = (u64)-1;
4173
4174         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4175                 page = list_entry(pages->prev, struct page, lru);
4176
4177                 prefetchw(&page->flags);
4178                 list_del(&page->lru);
4179                 if (add_to_page_cache_lru(page, mapping,
4180                                         page->index,
4181                                         readahead_gfp_mask(mapping))) {
4182                         put_page(page);
4183                         continue;
4184                 }
4185
4186                 pagepool[nr++] = page;
4187                 if (nr < ARRAY_SIZE(pagepool))
4188                         continue;
4189                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4190                                    &bio, 0, &bio_flags, &prev_em_start);
4191                 nr = 0;
4192         }
4193         if (nr)
4194                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4195                                    &bio, 0, &bio_flags, &prev_em_start);
4196
4197         if (em_cached)
4198                 free_extent_map(em_cached);
4199
4200         BUG_ON(!list_empty(pages));
4201         if (bio)
4202                 return submit_one_bio(bio, 0, bio_flags);
4203         return 0;
4204 }
4205
4206 /*
4207  * basic invalidatepage code, this waits on any locked or writeback
4208  * ranges corresponding to the page, and then deletes any extent state
4209  * records from the tree
4210  */
4211 int extent_invalidatepage(struct extent_io_tree *tree,
4212                           struct page *page, unsigned long offset)
4213 {
4214         struct extent_state *cached_state = NULL;
4215         u64 start = page_offset(page);
4216         u64 end = start + PAGE_SIZE - 1;
4217         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4218
4219         start += ALIGN(offset, blocksize);
4220         if (start > end)
4221                 return 0;
4222
4223         lock_extent_bits(tree, start, end, &cached_state);
4224         wait_on_page_writeback(page);
4225         clear_extent_bit(tree, start, end,
4226                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4227                          EXTENT_DO_ACCOUNTING,
4228                          1, 1, &cached_state, GFP_NOFS);
4229         return 0;
4230 }
4231
4232 /*
4233  * a helper for releasepage, this tests for areas of the page that
4234  * are locked or under IO and drops the related state bits if it is safe
4235  * to drop the page.
4236  */
4237 static int try_release_extent_state(struct extent_map_tree *map,
4238                                     struct extent_io_tree *tree,
4239                                     struct page *page, gfp_t mask)
4240 {
4241         u64 start = page_offset(page);
4242         u64 end = start + PAGE_SIZE - 1;
4243         int ret = 1;
4244
4245         if (test_range_bit(tree, start, end,
4246                            EXTENT_IOBITS, 0, NULL))
4247                 ret = 0;
4248         else {
4249                 /*
4250                  * at this point we can safely clear everything except the
4251                  * locked bit and the nodatasum bit
4252                  */
4253                 ret = clear_extent_bit(tree, start, end,
4254                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4255                                  0, 0, NULL, mask);
4256
4257                 /* if clear_extent_bit failed for enomem reasons,
4258                  * we can't allow the release to continue.
4259                  */
4260                 if (ret < 0)
4261                         ret = 0;
4262                 else
4263                         ret = 1;
4264         }
4265         return ret;
4266 }
4267
4268 /*
4269  * a helper for releasepage.  As long as there are no locked extents
4270  * in the range corresponding to the page, both state records and extent
4271  * map records are removed
4272  */
4273 int try_release_extent_mapping(struct extent_map_tree *map,
4274                                struct extent_io_tree *tree, struct page *page,
4275                                gfp_t mask)
4276 {
4277         struct extent_map *em;
4278         u64 start = page_offset(page);
4279         u64 end = start + PAGE_SIZE - 1;
4280
4281         if (gfpflags_allow_blocking(mask) &&
4282             page->mapping->host->i_size > SZ_16M) {
4283                 u64 len;
4284                 while (start <= end) {
4285                         len = end - start + 1;
4286                         write_lock(&map->lock);
4287                         em = lookup_extent_mapping(map, start, len);
4288                         if (!em) {
4289                                 write_unlock(&map->lock);
4290                                 break;
4291                         }
4292                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4293                             em->start != start) {
4294                                 write_unlock(&map->lock);
4295                                 free_extent_map(em);
4296                                 break;
4297                         }
4298                         if (!test_range_bit(tree, em->start,
4299                                             extent_map_end(em) - 1,
4300                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4301                                             0, NULL)) {
4302                                 remove_extent_mapping(map, em);
4303                                 /* once for the rb tree */
4304                                 free_extent_map(em);
4305                         }
4306                         start = extent_map_end(em);
4307                         write_unlock(&map->lock);
4308
4309                         /* once for us */
4310                         free_extent_map(em);
4311                 }
4312         }
4313         return try_release_extent_state(map, tree, page, mask);
4314 }
4315
4316 /*
4317  * helper function for fiemap, which doesn't want to see any holes.
4318  * This maps until we find something past 'last'
4319  */
4320 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4321                                                 u64 offset,
4322                                                 u64 last,
4323                                                 get_extent_t *get_extent)
4324 {
4325         u64 sectorsize = btrfs_inode_sectorsize(inode);
4326         struct extent_map *em;
4327         u64 len;
4328
4329         if (offset >= last)
4330                 return NULL;
4331
4332         while (1) {
4333                 len = last - offset;
4334                 if (len == 0)
4335                         break;
4336                 len = ALIGN(len, sectorsize);
4337                 em = get_extent(BTRFS_I(inode), NULL, 0, offset, len, 0);
4338                 if (IS_ERR_OR_NULL(em))
4339                         return em;
4340
4341                 /* if this isn't a hole return it */
4342                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4343                     em->block_start != EXTENT_MAP_HOLE) {
4344                         return em;
4345                 }
4346
4347                 /* this is a hole, advance to the next extent */
4348                 offset = extent_map_end(em);
4349                 free_extent_map(em);
4350                 if (offset >= last)
4351                         break;
4352         }
4353         return NULL;
4354 }
4355
4356 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4357                 __u64 start, __u64 len, get_extent_t *get_extent)
4358 {
4359         int ret = 0;
4360         u64 off = start;
4361         u64 max = start + len;
4362         u32 flags = 0;
4363         u32 found_type;
4364         u64 last;
4365         u64 last_for_get_extent = 0;
4366         u64 disko = 0;
4367         u64 isize = i_size_read(inode);
4368         struct btrfs_key found_key;
4369         struct extent_map *em = NULL;
4370         struct extent_state *cached_state = NULL;
4371         struct btrfs_path *path;
4372         struct btrfs_root *root = BTRFS_I(inode)->root;
4373         int end = 0;
4374         u64 em_start = 0;
4375         u64 em_len = 0;
4376         u64 em_end = 0;
4377
4378         if (len == 0)
4379                 return -EINVAL;
4380
4381         path = btrfs_alloc_path();
4382         if (!path)
4383                 return -ENOMEM;
4384         path->leave_spinning = 1;
4385
4386         start = round_down(start, btrfs_inode_sectorsize(inode));
4387         len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
4388
4389         /*
4390          * lookup the last file extent.  We're not using i_size here
4391          * because there might be preallocation past i_size
4392          */
4393         ret = btrfs_lookup_file_extent(NULL, root, path,
4394                         btrfs_ino(BTRFS_I(inode)), -1, 0);
4395         if (ret < 0) {
4396                 btrfs_free_path(path);
4397                 return ret;
4398         } else {
4399                 WARN_ON(!ret);
4400                 if (ret == 1)
4401                         ret = 0;
4402         }
4403
4404         path->slots[0]--;
4405         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4406         found_type = found_key.type;
4407
4408         /* No extents, but there might be delalloc bits */
4409         if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) ||
4410             found_type != BTRFS_EXTENT_DATA_KEY) {
4411                 /* have to trust i_size as the end */
4412                 last = (u64)-1;
4413                 last_for_get_extent = isize;
4414         } else {
4415                 /*
4416                  * remember the start of the last extent.  There are a
4417                  * bunch of different factors that go into the length of the
4418                  * extent, so its much less complex to remember where it started
4419                  */
4420                 last = found_key.offset;
4421                 last_for_get_extent = last + 1;
4422         }
4423         btrfs_release_path(path);
4424
4425         /*
4426          * we might have some extents allocated but more delalloc past those
4427          * extents.  so, we trust isize unless the start of the last extent is
4428          * beyond isize
4429          */
4430         if (last < isize) {
4431                 last = (u64)-1;
4432                 last_for_get_extent = isize;
4433         }
4434
4435         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4436                          &cached_state);
4437
4438         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4439                                    get_extent);
4440         if (!em)
4441                 goto out;
4442         if (IS_ERR(em)) {
4443                 ret = PTR_ERR(em);
4444                 goto out;
4445         }
4446
4447         while (!end) {
4448                 u64 offset_in_extent = 0;
4449
4450                 /* break if the extent we found is outside the range */
4451                 if (em->start >= max || extent_map_end(em) < off)
4452                         break;
4453
4454                 /*
4455                  * get_extent may return an extent that starts before our
4456                  * requested range.  We have to make sure the ranges
4457                  * we return to fiemap always move forward and don't
4458                  * overlap, so adjust the offsets here
4459                  */
4460                 em_start = max(em->start, off);
4461
4462                 /*
4463                  * record the offset from the start of the extent
4464                  * for adjusting the disk offset below.  Only do this if the
4465                  * extent isn't compressed since our in ram offset may be past
4466                  * what we have actually allocated on disk.
4467                  */
4468                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4469                         offset_in_extent = em_start - em->start;
4470                 em_end = extent_map_end(em);
4471                 em_len = em_end - em_start;
4472                 disko = 0;
4473                 flags = 0;
4474
4475                 /*
4476                  * bump off for our next call to get_extent
4477                  */
4478                 off = extent_map_end(em);
4479                 if (off >= max)
4480                         end = 1;
4481
4482                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4483                         end = 1;
4484                         flags |= FIEMAP_EXTENT_LAST;
4485                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4486                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4487                                   FIEMAP_EXTENT_NOT_ALIGNED);
4488                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4489                         flags |= (FIEMAP_EXTENT_DELALLOC |
4490                                   FIEMAP_EXTENT_UNKNOWN);
4491                 } else if (fieinfo->fi_extents_max) {
4492                         struct btrfs_trans_handle *trans;
4493
4494                         u64 bytenr = em->block_start -
4495                                 (em->start - em->orig_start);
4496
4497                         disko = em->block_start + offset_in_extent;
4498
4499                         /*
4500                          * We need a trans handle to get delayed refs
4501                          */
4502                         trans = btrfs_join_transaction(root);
4503                         /*
4504                          * It's OK if we can't start a trans we can still check
4505                          * from commit_root
4506                          */
4507                         if (IS_ERR(trans))
4508                                 trans = NULL;
4509
4510                         /*
4511                          * As btrfs supports shared space, this information
4512                          * can be exported to userspace tools via
4513                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4514                          * then we're just getting a count and we can skip the
4515                          * lookup stuff.
4516                          */
4517                         ret = btrfs_check_shared(trans, root->fs_info,
4518                                         root->objectid,
4519                                         btrfs_ino(BTRFS_I(inode)), bytenr);
4520                         if (trans)
4521                                 btrfs_end_transaction(trans);
4522                         if (ret < 0)
4523                                 goto out_free;
4524                         if (ret)
4525                                 flags |= FIEMAP_EXTENT_SHARED;
4526                         ret = 0;
4527                 }
4528                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4529                         flags |= FIEMAP_EXTENT_ENCODED;
4530                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4531                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4532
4533                 free_extent_map(em);
4534                 em = NULL;
4535                 if ((em_start >= last) || em_len == (u64)-1 ||
4536                    (last == (u64)-1 && isize <= em_end)) {
4537                         flags |= FIEMAP_EXTENT_LAST;
4538                         end = 1;
4539                 }
4540
4541                 /* now scan forward to see if this is really the last extent. */
4542                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4543                                            get_extent);
4544                 if (IS_ERR(em)) {
4545                         ret = PTR_ERR(em);
4546                         goto out;
4547                 }
4548                 if (!em) {
4549                         flags |= FIEMAP_EXTENT_LAST;
4550                         end = 1;
4551                 }
4552                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4553                                               em_len, flags);
4554                 if (ret) {
4555                         if (ret == 1)
4556                                 ret = 0;
4557                         goto out_free;
4558                 }
4559         }
4560 out_free:
4561         free_extent_map(em);
4562 out:
4563         btrfs_free_path(path);
4564         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4565                              &cached_state, GFP_NOFS);
4566         return ret;
4567 }
4568
4569 static void __free_extent_buffer(struct extent_buffer *eb)
4570 {
4571         btrfs_leak_debug_del(&eb->leak_list);
4572         kmem_cache_free(extent_buffer_cache, eb);
4573 }
4574
4575 int extent_buffer_under_io(struct extent_buffer *eb)
4576 {
4577         return (atomic_read(&eb->io_pages) ||
4578                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4579                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4580 }
4581
4582 /*
4583  * Helper for releasing extent buffer page.
4584  */
4585 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4586 {
4587         unsigned long index;
4588         struct page *page;
4589         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4590
4591         BUG_ON(extent_buffer_under_io(eb));
4592
4593         index = num_extent_pages(eb->start, eb->len);
4594         if (index == 0)
4595                 return;
4596
4597         do {
4598                 index--;
4599                 page = eb->pages[index];
4600                 if (!page)
4601                         continue;
4602                 if (mapped)
4603                         spin_lock(&page->mapping->private_lock);
4604                 /*
4605                  * We do this since we'll remove the pages after we've
4606                  * removed the eb from the radix tree, so we could race
4607                  * and have this page now attached to the new eb.  So
4608                  * only clear page_private if it's still connected to
4609                  * this eb.
4610                  */
4611                 if (PagePrivate(page) &&
4612                     page->private == (unsigned long)eb) {
4613                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4614                         BUG_ON(PageDirty(page));
4615                         BUG_ON(PageWriteback(page));
4616                         /*
4617                          * We need to make sure we haven't be attached
4618                          * to a new eb.
4619                          */
4620                         ClearPagePrivate(page);
4621                         set_page_private(page, 0);
4622                         /* One for the page private */
4623                         put_page(page);
4624                 }
4625
4626                 if (mapped)
4627                         spin_unlock(&page->mapping->private_lock);
4628
4629                 /* One for when we allocated the page */
4630                 put_page(page);
4631         } while (index != 0);
4632 }
4633
4634 /*
4635  * Helper for releasing the extent buffer.
4636  */
4637 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4638 {
4639         btrfs_release_extent_buffer_page(eb);
4640         __free_extent_buffer(eb);
4641 }
4642
4643 static struct extent_buffer *
4644 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4645                       unsigned long len)
4646 {
4647         struct extent_buffer *eb = NULL;
4648
4649         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4650         eb->start = start;
4651         eb->len = len;
4652         eb->fs_info = fs_info;
4653         eb->bflags = 0;
4654         rwlock_init(&eb->lock);
4655         atomic_set(&eb->write_locks, 0);
4656         atomic_set(&eb->read_locks, 0);
4657         atomic_set(&eb->blocking_readers, 0);
4658         atomic_set(&eb->blocking_writers, 0);
4659         atomic_set(&eb->spinning_readers, 0);
4660         atomic_set(&eb->spinning_writers, 0);
4661         eb->lock_nested = 0;
4662         init_waitqueue_head(&eb->write_lock_wq);
4663         init_waitqueue_head(&eb->read_lock_wq);
4664
4665         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4666
4667         spin_lock_init(&eb->refs_lock);
4668         atomic_set(&eb->refs, 1);
4669         atomic_set(&eb->io_pages, 0);
4670
4671         /*
4672          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4673          */
4674         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4675                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4676         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4677
4678         return eb;
4679 }
4680
4681 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4682 {
4683         unsigned long i;
4684         struct page *p;
4685         struct extent_buffer *new;
4686         unsigned long num_pages = num_extent_pages(src->start, src->len);
4687
4688         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4689         if (new == NULL)
4690                 return NULL;
4691
4692         for (i = 0; i < num_pages; i++) {
4693                 p = alloc_page(GFP_NOFS);
4694                 if (!p) {
4695                         btrfs_release_extent_buffer(new);
4696                         return NULL;
4697                 }
4698                 attach_extent_buffer_page(new, p);
4699                 WARN_ON(PageDirty(p));
4700                 SetPageUptodate(p);
4701                 new->pages[i] = p;
4702                 copy_page(page_address(p), page_address(src->pages[i]));
4703         }
4704
4705         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4706         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4707
4708         return new;
4709 }
4710
4711 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4712                                                   u64 start, unsigned long len)
4713 {
4714         struct extent_buffer *eb;
4715         unsigned long num_pages;
4716         unsigned long i;
4717
4718         num_pages = num_extent_pages(start, len);
4719
4720         eb = __alloc_extent_buffer(fs_info, start, len);
4721         if (!eb)
4722                 return NULL;
4723
4724         for (i = 0; i < num_pages; i++) {
4725                 eb->pages[i] = alloc_page(GFP_NOFS);
4726                 if (!eb->pages[i])
4727                         goto err;
4728         }
4729         set_extent_buffer_uptodate(eb);
4730         btrfs_set_header_nritems(eb, 0);
4731         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4732
4733         return eb;
4734 err:
4735         for (; i > 0; i--)
4736                 __free_page(eb->pages[i - 1]);
4737         __free_extent_buffer(eb);
4738         return NULL;
4739 }
4740
4741 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4742                                                 u64 start)
4743 {
4744         return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
4745 }
4746
4747 static void check_buffer_tree_ref(struct extent_buffer *eb)
4748 {
4749         int refs;
4750         /* the ref bit is tricky.  We have to make sure it is set
4751          * if we have the buffer dirty.   Otherwise the
4752          * code to free a buffer can end up dropping a dirty
4753          * page
4754          *
4755          * Once the ref bit is set, it won't go away while the
4756          * buffer is dirty or in writeback, and it also won't
4757          * go away while we have the reference count on the
4758          * eb bumped.
4759          *
4760          * We can't just set the ref bit without bumping the
4761          * ref on the eb because free_extent_buffer might
4762          * see the ref bit and try to clear it.  If this happens
4763          * free_extent_buffer might end up dropping our original
4764          * ref by mistake and freeing the page before we are able
4765          * to add one more ref.
4766          *
4767          * So bump the ref count first, then set the bit.  If someone
4768          * beat us to it, drop the ref we added.
4769          */
4770         refs = atomic_read(&eb->refs);
4771         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4772                 return;
4773
4774         spin_lock(&eb->refs_lock);
4775         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4776                 atomic_inc(&eb->refs);
4777         spin_unlock(&eb->refs_lock);
4778 }
4779
4780 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4781                 struct page *accessed)
4782 {
4783         unsigned long num_pages, i;
4784
4785         check_buffer_tree_ref(eb);
4786
4787         num_pages = num_extent_pages(eb->start, eb->len);
4788         for (i = 0; i < num_pages; i++) {
4789                 struct page *p = eb->pages[i];
4790
4791                 if (p != accessed)
4792                         mark_page_accessed(p);
4793         }
4794 }
4795
4796 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4797                                          u64 start)
4798 {
4799         struct extent_buffer *eb;
4800
4801         rcu_read_lock();
4802         eb = radix_tree_lookup(&fs_info->buffer_radix,
4803                                start >> PAGE_SHIFT);
4804         if (eb && atomic_inc_not_zero(&eb->refs)) {
4805                 rcu_read_unlock();
4806                 /*
4807                  * Lock our eb's refs_lock to avoid races with
4808                  * free_extent_buffer. When we get our eb it might be flagged
4809                  * with EXTENT_BUFFER_STALE and another task running
4810                  * free_extent_buffer might have seen that flag set,
4811                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4812                  * writeback flags not set) and it's still in the tree (flag
4813                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4814                  * of decrementing the extent buffer's reference count twice.
4815                  * So here we could race and increment the eb's reference count,
4816                  * clear its stale flag, mark it as dirty and drop our reference
4817                  * before the other task finishes executing free_extent_buffer,
4818                  * which would later result in an attempt to free an extent
4819                  * buffer that is dirty.
4820                  */
4821                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4822                         spin_lock(&eb->refs_lock);
4823                         spin_unlock(&eb->refs_lock);
4824                 }
4825                 mark_extent_buffer_accessed(eb, NULL);
4826                 return eb;
4827         }
4828         rcu_read_unlock();
4829
4830         return NULL;
4831 }
4832
4833 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4834 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4835                                         u64 start)
4836 {
4837         struct extent_buffer *eb, *exists = NULL;
4838         int ret;
4839
4840         eb = find_extent_buffer(fs_info, start);
4841         if (eb)
4842                 return eb;
4843         eb = alloc_dummy_extent_buffer(fs_info, start);
4844         if (!eb)
4845                 return NULL;
4846         eb->fs_info = fs_info;
4847 again:
4848         ret = radix_tree_preload(GFP_NOFS);
4849         if (ret)
4850                 goto free_eb;
4851         spin_lock(&fs_info->buffer_lock);
4852         ret = radix_tree_insert(&fs_info->buffer_radix,
4853                                 start >> PAGE_SHIFT, eb);
4854         spin_unlock(&fs_info->buffer_lock);
4855         radix_tree_preload_end();
4856         if (ret == -EEXIST) {
4857                 exists = find_extent_buffer(fs_info, start);
4858                 if (exists)
4859                         goto free_eb;
4860                 else
4861                         goto again;
4862         }
4863         check_buffer_tree_ref(eb);
4864         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4865
4866         /*
4867          * We will free dummy extent buffer's if they come into
4868          * free_extent_buffer with a ref count of 2, but if we are using this we
4869          * want the buffers to stay in memory until we're done with them, so
4870          * bump the ref count again.
4871          */
4872         atomic_inc(&eb->refs);
4873         return eb;
4874 free_eb:
4875         btrfs_release_extent_buffer(eb);
4876         return exists;
4877 }
4878 #endif
4879
4880 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4881                                           u64 start)
4882 {
4883         unsigned long len = fs_info->nodesize;
4884         unsigned long num_pages = num_extent_pages(start, len);
4885         unsigned long i;
4886         unsigned long index = start >> PAGE_SHIFT;
4887         struct extent_buffer *eb;
4888         struct extent_buffer *exists = NULL;
4889         struct page *p;
4890         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4891         int uptodate = 1;
4892         int ret;
4893
4894         if (!IS_ALIGNED(start, fs_info->sectorsize)) {
4895                 btrfs_err(fs_info, "bad tree block start %llu", start);
4896                 return ERR_PTR(-EINVAL);
4897         }
4898
4899         eb = find_extent_buffer(fs_info, start);
4900         if (eb)
4901                 return eb;
4902
4903         eb = __alloc_extent_buffer(fs_info, start, len);
4904         if (!eb)
4905                 return ERR_PTR(-ENOMEM);
4906
4907         for (i = 0; i < num_pages; i++, index++) {
4908                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4909                 if (!p) {
4910                         exists = ERR_PTR(-ENOMEM);
4911                         goto free_eb;
4912                 }
4913
4914                 spin_lock(&mapping->private_lock);
4915                 if (PagePrivate(p)) {
4916                         /*
4917                          * We could have already allocated an eb for this page
4918                          * and attached one so lets see if we can get a ref on
4919                          * the existing eb, and if we can we know it's good and
4920                          * we can just return that one, else we know we can just
4921                          * overwrite page->private.
4922                          */
4923                         exists = (struct extent_buffer *)p->private;
4924                         if (atomic_inc_not_zero(&exists->refs)) {
4925                                 spin_unlock(&mapping->private_lock);
4926                                 unlock_page(p);
4927                                 put_page(p);
4928                                 mark_extent_buffer_accessed(exists, p);
4929                                 goto free_eb;
4930                         }
4931                         exists = NULL;
4932
4933                         /*
4934                          * Do this so attach doesn't complain and we need to
4935                          * drop the ref the old guy had.
4936                          */
4937                         ClearPagePrivate(p);
4938                         WARN_ON(PageDirty(p));
4939                         put_page(p);
4940                 }
4941                 attach_extent_buffer_page(eb, p);
4942                 spin_unlock(&mapping->private_lock);
4943                 WARN_ON(PageDirty(p));
4944                 eb->pages[i] = p;
4945                 if (!PageUptodate(p))
4946                         uptodate = 0;
4947
4948                 /*
4949                  * see below about how we avoid a nasty race with release page
4950                  * and why we unlock later
4951                  */
4952         }
4953         if (uptodate)
4954                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4955 again:
4956         ret = radix_tree_preload(GFP_NOFS);
4957         if (ret) {
4958                 exists = ERR_PTR(ret);
4959                 goto free_eb;
4960         }
4961
4962         spin_lock(&fs_info->buffer_lock);
4963         ret = radix_tree_insert(&fs_info->buffer_radix,
4964                                 start >> PAGE_SHIFT, eb);
4965         spin_unlock(&fs_info->buffer_lock);
4966         radix_tree_preload_end();
4967         if (ret == -EEXIST) {
4968                 exists = find_extent_buffer(fs_info, start);
4969                 if (exists)
4970                         goto free_eb;
4971                 else
4972                         goto again;
4973         }
4974         /* add one reference for the tree */
4975         check_buffer_tree_ref(eb);
4976         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4977
4978         /*
4979          * there is a race where release page may have
4980          * tried to find this extent buffer in the radix
4981          * but failed.  It will tell the VM it is safe to
4982          * reclaim the, and it will clear the page private bit.
4983          * We must make sure to set the page private bit properly
4984          * after the extent buffer is in the radix tree so
4985          * it doesn't get lost
4986          */
4987         SetPageChecked(eb->pages[0]);
4988         for (i = 1; i < num_pages; i++) {
4989                 p = eb->pages[i];
4990                 ClearPageChecked(p);
4991                 unlock_page(p);
4992         }
4993         unlock_page(eb->pages[0]);
4994         return eb;
4995
4996 free_eb:
4997         WARN_ON(!atomic_dec_and_test(&eb->refs));
4998         for (i = 0; i < num_pages; i++) {
4999                 if (eb->pages[i])
5000                         unlock_page(eb->pages[i]);
5001         }
5002
5003         btrfs_release_extent_buffer(eb);
5004         return exists;
5005 }
5006
5007 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5008 {
5009         struct extent_buffer *eb =
5010                         container_of(head, struct extent_buffer, rcu_head);
5011
5012         __free_extent_buffer(eb);
5013 }
5014
5015 /* Expects to have eb->eb_lock already held */
5016 static int release_extent_buffer(struct extent_buffer *eb)
5017 {
5018         WARN_ON(atomic_read(&eb->refs) == 0);
5019         if (atomic_dec_and_test(&eb->refs)) {
5020                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5021                         struct btrfs_fs_info *fs_info = eb->fs_info;
5022
5023                         spin_unlock(&eb->refs_lock);
5024
5025                         spin_lock(&fs_info->buffer_lock);
5026                         radix_tree_delete(&fs_info->buffer_radix,
5027                                           eb->start >> PAGE_SHIFT);
5028                         spin_unlock(&fs_info->buffer_lock);
5029                 } else {
5030                         spin_unlock(&eb->refs_lock);
5031                 }
5032
5033                 /* Should be safe to release our pages at this point */
5034                 btrfs_release_extent_buffer_page(eb);
5035 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5036                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5037                         __free_extent_buffer(eb);
5038                         return 1;
5039                 }
5040 #endif
5041                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5042                 return 1;
5043         }
5044         spin_unlock(&eb->refs_lock);
5045
5046         return 0;
5047 }
5048
5049 void free_extent_buffer(struct extent_buffer *eb)
5050 {
5051         int refs;
5052         int old;
5053         if (!eb)
5054                 return;
5055
5056         while (1) {
5057                 refs = atomic_read(&eb->refs);
5058                 if (refs <= 3)
5059                         break;
5060                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5061                 if (old == refs)
5062                         return;
5063         }
5064
5065         spin_lock(&eb->refs_lock);
5066         if (atomic_read(&eb->refs) == 2 &&
5067             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5068                 atomic_dec(&eb->refs);
5069
5070         if (atomic_read(&eb->refs) == 2 &&
5071             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5072             !extent_buffer_under_io(eb) &&
5073             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5074                 atomic_dec(&eb->refs);
5075
5076         /*
5077          * I know this is terrible, but it's temporary until we stop tracking
5078          * the uptodate bits and such for the extent buffers.
5079          */
5080         release_extent_buffer(eb);
5081 }
5082
5083 void free_extent_buffer_stale(struct extent_buffer *eb)
5084 {
5085         if (!eb)
5086                 return;
5087
5088         spin_lock(&eb->refs_lock);
5089         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5090
5091         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5092             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5093                 atomic_dec(&eb->refs);
5094         release_extent_buffer(eb);
5095 }
5096
5097 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5098 {
5099         unsigned long i;
5100         unsigned long num_pages;
5101         struct page *page;
5102
5103         num_pages = num_extent_pages(eb->start, eb->len);
5104
5105         for (i = 0; i < num_pages; i++) {
5106                 page = eb->pages[i];
5107                 if (!PageDirty(page))
5108                         continue;
5109
5110                 lock_page(page);
5111                 WARN_ON(!PagePrivate(page));
5112
5113                 clear_page_dirty_for_io(page);
5114                 spin_lock_irq(&page->mapping->tree_lock);
5115                 if (!PageDirty(page)) {
5116                         radix_tree_tag_clear(&page->mapping->page_tree,
5117                                                 page_index(page),
5118                                                 PAGECACHE_TAG_DIRTY);
5119                 }
5120                 spin_unlock_irq(&page->mapping->tree_lock);
5121                 ClearPageError(page);
5122                 unlock_page(page);
5123         }
5124         WARN_ON(atomic_read(&eb->refs) == 0);
5125 }
5126
5127 int set_extent_buffer_dirty(struct extent_buffer *eb)
5128 {
5129         unsigned long i;
5130         unsigned long num_pages;
5131         int was_dirty = 0;
5132
5133         check_buffer_tree_ref(eb);
5134
5135         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5136
5137         num_pages = num_extent_pages(eb->start, eb->len);
5138         WARN_ON(atomic_read(&eb->refs) == 0);
5139         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5140
5141         for (i = 0; i < num_pages; i++)
5142                 set_page_dirty(eb->pages[i]);
5143         return was_dirty;
5144 }
5145
5146 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5147 {
5148         unsigned long i;
5149         struct page *page;
5150         unsigned long num_pages;
5151
5152         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5153         num_pages = num_extent_pages(eb->start, eb->len);
5154         for (i = 0; i < num_pages; i++) {
5155                 page = eb->pages[i];
5156                 if (page)
5157                         ClearPageUptodate(page);
5158         }
5159 }
5160
5161 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5162 {
5163         unsigned long i;
5164         struct page *page;
5165         unsigned long num_pages;
5166
5167         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5168         num_pages = num_extent_pages(eb->start, eb->len);
5169         for (i = 0; i < num_pages; i++) {
5170                 page = eb->pages[i];
5171                 SetPageUptodate(page);
5172         }
5173 }
5174
5175 int extent_buffer_uptodate(struct extent_buffer *eb)
5176 {
5177         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5178 }
5179
5180 int read_extent_buffer_pages(struct extent_io_tree *tree,
5181                              struct extent_buffer *eb, int wait,
5182                              get_extent_t *get_extent, int mirror_num)
5183 {
5184         unsigned long i;
5185         struct page *page;
5186         int err;
5187         int ret = 0;
5188         int locked_pages = 0;
5189         int all_uptodate = 1;
5190         unsigned long num_pages;
5191         unsigned long num_reads = 0;
5192         struct bio *bio = NULL;
5193         unsigned long bio_flags = 0;
5194
5195         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5196                 return 0;
5197
5198         num_pages = num_extent_pages(eb->start, eb->len);
5199         for (i = 0; i < num_pages; i++) {
5200                 page = eb->pages[i];
5201                 if (wait == WAIT_NONE) {
5202                         if (!trylock_page(page))
5203                                 goto unlock_exit;
5204                 } else {
5205                         lock_page(page);
5206                 }
5207                 locked_pages++;
5208         }
5209         /*
5210          * We need to firstly lock all pages to make sure that
5211          * the uptodate bit of our pages won't be affected by
5212          * clear_extent_buffer_uptodate().
5213          */
5214         for (i = 0; i < num_pages; i++) {
5215                 page = eb->pages[i];
5216                 if (!PageUptodate(page)) {
5217                         num_reads++;
5218                         all_uptodate = 0;
5219                 }
5220         }
5221
5222         if (all_uptodate) {
5223                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5224                 goto unlock_exit;
5225         }
5226
5227         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5228         eb->read_mirror = 0;
5229         atomic_set(&eb->io_pages, num_reads);
5230         for (i = 0; i < num_pages; i++) {
5231                 page = eb->pages[i];
5232
5233                 if (!PageUptodate(page)) {
5234                         if (ret) {
5235                                 atomic_dec(&eb->io_pages);
5236                                 unlock_page(page);
5237                                 continue;
5238                         }
5239
5240                         ClearPageError(page);
5241                         err = __extent_read_full_page(tree, page,
5242                                                       get_extent, &bio,
5243                                                       mirror_num, &bio_flags,
5244                                                       REQ_META);
5245                         if (err) {
5246                                 ret = err;
5247                                 /*
5248                                  * We use &bio in above __extent_read_full_page,
5249                                  * so we ensure that if it returns error, the
5250                                  * current page fails to add itself to bio and
5251                                  * it's been unlocked.
5252                                  *
5253                                  * We must dec io_pages by ourselves.
5254                                  */
5255                                 atomic_dec(&eb->io_pages);
5256                         }
5257                 } else {
5258                         unlock_page(page);
5259                 }
5260         }
5261
5262         if (bio) {
5263                 err = submit_one_bio(bio, mirror_num, bio_flags);
5264                 if (err)
5265                         return err;
5266         }
5267
5268         if (ret || wait != WAIT_COMPLETE)
5269                 return ret;
5270
5271         for (i = 0; i < num_pages; i++) {
5272                 page = eb->pages[i];
5273                 wait_on_page_locked(page);
5274                 if (!PageUptodate(page))
5275                         ret = -EIO;
5276         }
5277
5278         return ret;
5279
5280 unlock_exit:
5281         while (locked_pages > 0) {
5282                 locked_pages--;
5283                 page = eb->pages[locked_pages];
5284                 unlock_page(page);
5285         }
5286         return ret;
5287 }
5288
5289 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5290                         unsigned long start,
5291                         unsigned long len)
5292 {
5293         size_t cur;
5294         size_t offset;
5295         struct page *page;
5296         char *kaddr;
5297         char *dst = (char *)dstv;
5298         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5299         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5300
5301         WARN_ON(start > eb->len);
5302         WARN_ON(start + len > eb->start + eb->len);
5303
5304         offset = (start_offset + start) & (PAGE_SIZE - 1);
5305
5306         while (len > 0) {
5307                 page = eb->pages[i];
5308
5309                 cur = min(len, (PAGE_SIZE - offset));
5310                 kaddr = page_address(page);
5311                 memcpy(dst, kaddr + offset, cur);
5312
5313                 dst += cur;
5314                 len -= cur;
5315                 offset = 0;
5316                 i++;
5317         }
5318 }
5319
5320 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5321                         unsigned long start,
5322                         unsigned long len)
5323 {
5324         size_t cur;
5325         size_t offset;
5326         struct page *page;
5327         char *kaddr;
5328         char __user *dst = (char __user *)dstv;
5329         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5330         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5331         int ret = 0;
5332
5333         WARN_ON(start > eb->len);
5334         WARN_ON(start + len > eb->start + eb->len);
5335
5336         offset = (start_offset + start) & (PAGE_SIZE - 1);
5337
5338         while (len > 0) {
5339                 page = eb->pages[i];
5340
5341                 cur = min(len, (PAGE_SIZE - offset));
5342                 kaddr = page_address(page);
5343                 if (copy_to_user(dst, kaddr + offset, cur)) {
5344                         ret = -EFAULT;
5345                         break;
5346                 }
5347
5348                 dst += cur;
5349                 len -= cur;
5350                 offset = 0;
5351                 i++;
5352         }
5353
5354         return ret;
5355 }
5356
5357 /*
5358  * return 0 if the item is found within a page.
5359  * return 1 if the item spans two pages.
5360  * return -EINVAL otherwise.
5361  */
5362 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5363                                unsigned long min_len, char **map,
5364                                unsigned long *map_start,
5365                                unsigned long *map_len)
5366 {
5367         size_t offset = start & (PAGE_SIZE - 1);
5368         char *kaddr;
5369         struct page *p;
5370         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5371         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5372         unsigned long end_i = (start_offset + start + min_len - 1) >>
5373                 PAGE_SHIFT;
5374
5375         if (i != end_i)
5376                 return 1;
5377
5378         if (i == 0) {
5379                 offset = start_offset;
5380                 *map_start = 0;
5381         } else {
5382                 offset = 0;
5383                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5384         }
5385
5386         if (start + min_len > eb->len) {
5387                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5388                        eb->start, eb->len, start, min_len);
5389                 return -EINVAL;
5390         }
5391
5392         p = eb->pages[i];
5393         kaddr = page_address(p);
5394         *map = kaddr + offset;
5395         *map_len = PAGE_SIZE - offset;
5396         return 0;
5397 }
5398
5399 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5400                           unsigned long start,
5401                           unsigned long len)
5402 {
5403         size_t cur;
5404         size_t offset;
5405         struct page *page;
5406         char *kaddr;
5407         char *ptr = (char *)ptrv;
5408         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5409         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5410         int ret = 0;
5411
5412         WARN_ON(start > eb->len);
5413         WARN_ON(start + len > eb->start + eb->len);
5414
5415         offset = (start_offset + start) & (PAGE_SIZE - 1);
5416
5417         while (len > 0) {
5418                 page = eb->pages[i];
5419
5420                 cur = min(len, (PAGE_SIZE - offset));
5421
5422                 kaddr = page_address(page);
5423                 ret = memcmp(ptr, kaddr + offset, cur);
5424                 if (ret)
5425                         break;
5426
5427                 ptr += cur;
5428                 len -= cur;
5429                 offset = 0;
5430                 i++;
5431         }
5432         return ret;
5433 }
5434
5435 void write_extent_buffer_chunk_tree_uuid(struct extent_buffer *eb,
5436                 const void *srcv)
5437 {
5438         char *kaddr;
5439
5440         WARN_ON(!PageUptodate(eb->pages[0]));
5441         kaddr = page_address(eb->pages[0]);
5442         memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv,
5443                         BTRFS_FSID_SIZE);
5444 }
5445
5446 void write_extent_buffer_fsid(struct extent_buffer *eb, const void *srcv)
5447 {
5448         char *kaddr;
5449
5450         WARN_ON(!PageUptodate(eb->pages[0]));
5451         kaddr = page_address(eb->pages[0]);
5452         memcpy(kaddr + offsetof(struct btrfs_header, fsid), srcv,
5453                         BTRFS_FSID_SIZE);
5454 }
5455
5456 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5457                          unsigned long start, unsigned long len)
5458 {
5459         size_t cur;
5460         size_t offset;
5461         struct page *page;
5462         char *kaddr;
5463         char *src = (char *)srcv;
5464         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5465         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5466
5467         WARN_ON(start > eb->len);
5468         WARN_ON(start + len > eb->start + eb->len);
5469
5470         offset = (start_offset + start) & (PAGE_SIZE - 1);
5471
5472         while (len > 0) {
5473                 page = eb->pages[i];
5474                 WARN_ON(!PageUptodate(page));
5475
5476                 cur = min(len, PAGE_SIZE - offset);
5477                 kaddr = page_address(page);
5478                 memcpy(kaddr + offset, src, cur);
5479
5480                 src += cur;
5481                 len -= cur;
5482                 offset = 0;
5483                 i++;
5484         }
5485 }
5486
5487 void memzero_extent_buffer(struct extent_buffer *eb, unsigned long start,
5488                 unsigned long len)
5489 {
5490         size_t cur;
5491         size_t offset;
5492         struct page *page;
5493         char *kaddr;
5494         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5495         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5496
5497         WARN_ON(start > eb->len);
5498         WARN_ON(start + len > eb->start + eb->len);
5499
5500         offset = (start_offset + start) & (PAGE_SIZE - 1);
5501
5502         while (len > 0) {
5503                 page = eb->pages[i];
5504                 WARN_ON(!PageUptodate(page));
5505
5506                 cur = min(len, PAGE_SIZE - offset);
5507                 kaddr = page_address(page);
5508                 memset(kaddr + offset, 0, cur);
5509
5510                 len -= cur;
5511                 offset = 0;
5512                 i++;
5513         }
5514 }
5515
5516 void copy_extent_buffer_full(struct extent_buffer *dst,
5517                              struct extent_buffer *src)
5518 {
5519         int i;
5520         unsigned num_pages;
5521
5522         ASSERT(dst->len == src->len);
5523
5524         num_pages = num_extent_pages(dst->start, dst->len);
5525         for (i = 0; i < num_pages; i++)
5526                 copy_page(page_address(dst->pages[i]),
5527                                 page_address(src->pages[i]));
5528 }
5529
5530 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5531                         unsigned long dst_offset, unsigned long src_offset,
5532                         unsigned long len)
5533 {
5534         u64 dst_len = dst->len;
5535         size_t cur;
5536         size_t offset;
5537         struct page *page;
5538         char *kaddr;
5539         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5540         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5541
5542         WARN_ON(src->len != dst_len);
5543
5544         offset = (start_offset + dst_offset) &
5545                 (PAGE_SIZE - 1);
5546
5547         while (len > 0) {
5548                 page = dst->pages[i];
5549                 WARN_ON(!PageUptodate(page));
5550
5551                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5552
5553                 kaddr = page_address(page);
5554                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5555
5556                 src_offset += cur;
5557                 len -= cur;
5558                 offset = 0;
5559                 i++;
5560         }
5561 }
5562
5563 void le_bitmap_set(u8 *map, unsigned int start, int len)
5564 {
5565         u8 *p = map + BIT_BYTE(start);
5566         const unsigned int size = start + len;
5567         int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5568         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);
5569
5570         while (len - bits_to_set >= 0) {
5571                 *p |= mask_to_set;
5572                 len -= bits_to_set;
5573                 bits_to_set = BITS_PER_BYTE;
5574                 mask_to_set = ~0;
5575                 p++;
5576         }
5577         if (len) {
5578                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5579                 *p |= mask_to_set;
5580         }
5581 }
5582
5583 void le_bitmap_clear(u8 *map, unsigned int start, int len)
5584 {
5585         u8 *p = map + BIT_BYTE(start);
5586         const unsigned int size = start + len;
5587         int bits_to_clear = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5588         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(start);
5589
5590         while (len - bits_to_clear >= 0) {
5591                 *p &= ~mask_to_clear;
5592                 len -= bits_to_clear;
5593                 bits_to_clear = BITS_PER_BYTE;
5594                 mask_to_clear = ~0;
5595                 p++;
5596         }
5597         if (len) {
5598                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5599                 *p &= ~mask_to_clear;
5600         }
5601 }
5602
5603 /*
5604  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5605  * given bit number
5606  * @eb: the extent buffer
5607  * @start: offset of the bitmap item in the extent buffer
5608  * @nr: bit number
5609  * @page_index: return index of the page in the extent buffer that contains the
5610  * given bit number
5611  * @page_offset: return offset into the page given by page_index
5612  *
5613  * This helper hides the ugliness of finding the byte in an extent buffer which
5614  * contains a given bit.
5615  */
5616 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5617                                     unsigned long start, unsigned long nr,
5618                                     unsigned long *page_index,
5619                                     size_t *page_offset)
5620 {
5621         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5622         size_t byte_offset = BIT_BYTE(nr);
5623         size_t offset;
5624
5625         /*
5626          * The byte we want is the offset of the extent buffer + the offset of
5627          * the bitmap item in the extent buffer + the offset of the byte in the
5628          * bitmap item.
5629          */
5630         offset = start_offset + start + byte_offset;
5631
5632         *page_index = offset >> PAGE_SHIFT;
5633         *page_offset = offset & (PAGE_SIZE - 1);
5634 }
5635
5636 /**
5637  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5638  * @eb: the extent buffer
5639  * @start: offset of the bitmap item in the extent buffer
5640  * @nr: bit number to test
5641  */
5642 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5643                            unsigned long nr)
5644 {
5645         u8 *kaddr;
5646         struct page *page;
5647         unsigned long i;
5648         size_t offset;
5649
5650         eb_bitmap_offset(eb, start, nr, &i, &offset);
5651         page = eb->pages[i];
5652         WARN_ON(!PageUptodate(page));
5653         kaddr = page_address(page);
5654         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5655 }
5656
5657 /**
5658  * extent_buffer_bitmap_set - set an area of a bitmap
5659  * @eb: the extent buffer
5660  * @start: offset of the bitmap item in the extent buffer
5661  * @pos: bit number of the first bit
5662  * @len: number of bits to set
5663  */
5664 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5665                               unsigned long pos, unsigned long len)
5666 {
5667         u8 *kaddr;
5668         struct page *page;
5669         unsigned long i;
5670         size_t offset;
5671         const unsigned int size = pos + len;
5672         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5673         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5674
5675         eb_bitmap_offset(eb, start, pos, &i, &offset);
5676         page = eb->pages[i];
5677         WARN_ON(!PageUptodate(page));
5678         kaddr = page_address(page);
5679
5680         while (len >= bits_to_set) {
5681                 kaddr[offset] |= mask_to_set;
5682                 len -= bits_to_set;
5683                 bits_to_set = BITS_PER_BYTE;
5684                 mask_to_set = ~0;
5685                 if (++offset >= PAGE_SIZE && len > 0) {
5686                         offset = 0;
5687                         page = eb->pages[++i];
5688                         WARN_ON(!PageUptodate(page));
5689                         kaddr = page_address(page);
5690                 }
5691         }
5692         if (len) {
5693                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5694                 kaddr[offset] |= mask_to_set;
5695         }
5696 }
5697
5698
5699 /**
5700  * extent_buffer_bitmap_clear - clear an area of a bitmap
5701  * @eb: the extent buffer
5702  * @start: offset of the bitmap item in the extent buffer
5703  * @pos: bit number of the first bit
5704  * @len: number of bits to clear
5705  */
5706 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5707                                 unsigned long pos, unsigned long len)
5708 {
5709         u8 *kaddr;
5710         struct page *page;
5711         unsigned long i;
5712         size_t offset;
5713         const unsigned int size = pos + len;
5714         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5715         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5716
5717         eb_bitmap_offset(eb, start, pos, &i, &offset);
5718         page = eb->pages[i];
5719         WARN_ON(!PageUptodate(page));
5720         kaddr = page_address(page);
5721
5722         while (len >= bits_to_clear) {
5723                 kaddr[offset] &= ~mask_to_clear;
5724                 len -= bits_to_clear;
5725                 bits_to_clear = BITS_PER_BYTE;
5726                 mask_to_clear = ~0;
5727                 if (++offset >= PAGE_SIZE && len > 0) {
5728                         offset = 0;
5729                         page = eb->pages[++i];
5730                         WARN_ON(!PageUptodate(page));
5731                         kaddr = page_address(page);
5732                 }
5733         }
5734         if (len) {
5735                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5736                 kaddr[offset] &= ~mask_to_clear;
5737         }
5738 }
5739
5740 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5741 {
5742         unsigned long distance = (src > dst) ? src - dst : dst - src;
5743         return distance < len;
5744 }
5745
5746 static void copy_pages(struct page *dst_page, struct page *src_page,
5747                        unsigned long dst_off, unsigned long src_off,
5748                        unsigned long len)
5749 {
5750         char *dst_kaddr = page_address(dst_page);
5751         char *src_kaddr;
5752         int must_memmove = 0;
5753
5754         if (dst_page != src_page) {
5755                 src_kaddr = page_address(src_page);
5756         } else {
5757                 src_kaddr = dst_kaddr;
5758                 if (areas_overlap(src_off, dst_off, len))
5759                         must_memmove = 1;
5760         }
5761
5762         if (must_memmove)
5763                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5764         else
5765                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5766 }
5767
5768 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5769                            unsigned long src_offset, unsigned long len)
5770 {
5771         struct btrfs_fs_info *fs_info = dst->fs_info;
5772         size_t cur;
5773         size_t dst_off_in_page;
5774         size_t src_off_in_page;
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(fs_info,
5781                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5782                          src_offset, len, dst->len);
5783                 BUG_ON(1);
5784         }
5785         if (dst_offset + len > dst->len) {
5786                 btrfs_err(fs_info,
5787                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5788                          dst_offset, len, dst->len);
5789                 BUG_ON(1);
5790         }
5791
5792         while (len > 0) {
5793                 dst_off_in_page = (start_offset + dst_offset) &
5794                         (PAGE_SIZE - 1);
5795                 src_off_in_page = (start_offset + src_offset) &
5796                         (PAGE_SIZE - 1);
5797
5798                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5799                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5800
5801                 cur = min(len, (unsigned long)(PAGE_SIZE -
5802                                                src_off_in_page));
5803                 cur = min_t(unsigned long, cur,
5804                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5805
5806                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5807                            dst_off_in_page, src_off_in_page, cur);
5808
5809                 src_offset += cur;
5810                 dst_offset += cur;
5811                 len -= cur;
5812         }
5813 }
5814
5815 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5816                            unsigned long src_offset, unsigned long len)
5817 {
5818         struct btrfs_fs_info *fs_info = dst->fs_info;
5819         size_t cur;
5820         size_t dst_off_in_page;
5821         size_t src_off_in_page;
5822         unsigned long dst_end = dst_offset + len - 1;
5823         unsigned long src_end = src_offset + len - 1;
5824         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5825         unsigned long dst_i;
5826         unsigned long src_i;
5827
5828         if (src_offset + len > dst->len) {
5829                 btrfs_err(fs_info,
5830                           "memmove bogus src_offset %lu move len %lu len %lu",
5831                           src_offset, len, dst->len);
5832                 BUG_ON(1);
5833         }
5834         if (dst_offset + len > dst->len) {
5835                 btrfs_err(fs_info,
5836                           "memmove bogus dst_offset %lu move len %lu len %lu",
5837                           dst_offset, len, dst->len);
5838                 BUG_ON(1);
5839         }
5840         if (dst_offset < src_offset) {
5841                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5842                 return;
5843         }
5844         while (len > 0) {
5845                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5846                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5847
5848                 dst_off_in_page = (start_offset + dst_end) &
5849                         (PAGE_SIZE - 1);
5850                 src_off_in_page = (start_offset + src_end) &
5851                         (PAGE_SIZE - 1);
5852
5853                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5854                 cur = min(cur, dst_off_in_page + 1);
5855                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5856                            dst_off_in_page - cur + 1,
5857                            src_off_in_page - cur + 1, cur);
5858
5859                 dst_end -= cur;
5860                 src_end -= cur;
5861                 len -= cur;
5862         }
5863 }
5864
5865 int try_release_extent_buffer(struct page *page)
5866 {
5867         struct extent_buffer *eb;
5868
5869         /*
5870          * We need to make sure nobody is attaching this page to an eb right
5871          * now.
5872          */
5873         spin_lock(&page->mapping->private_lock);
5874         if (!PagePrivate(page)) {
5875                 spin_unlock(&page->mapping->private_lock);
5876                 return 1;
5877         }
5878
5879         eb = (struct extent_buffer *)page->private;
5880         BUG_ON(!eb);
5881
5882         /*
5883          * This is a little awful but should be ok, we need to make sure that
5884          * the eb doesn't disappear out from under us while we're looking at
5885          * this page.
5886          */
5887         spin_lock(&eb->refs_lock);
5888         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5889                 spin_unlock(&eb->refs_lock);
5890                 spin_unlock(&page->mapping->private_lock);
5891                 return 0;
5892         }
5893         spin_unlock(&page->mapping->private_lock);
5894
5895         /*
5896          * If tree ref isn't set then we know the ref on this eb is a real ref,
5897          * so just return, this page will likely be freed soon anyway.
5898          */
5899         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5900                 spin_unlock(&eb->refs_lock);
5901                 return 0;
5902         }
5903
5904         return release_extent_buffer(eb);
5905 }