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                        refcount_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         refcount_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 (refcount_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                                 refcount_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                         refcount_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                         refcount_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                         refcount_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                         err = -EAGAIN;
1718                         goto out;
1719                 }
1720
1721                 for (i = 0; i < ret; i++) {
1722                         if (page_ops & PAGE_SET_PRIVATE2)
1723                                 SetPagePrivate2(pages[i]);
1724
1725                         if (pages[i] == locked_page) {
1726                                 put_page(pages[i]);
1727                                 pages_locked++;
1728                                 continue;
1729                         }
1730                         if (page_ops & PAGE_CLEAR_DIRTY)
1731                                 clear_page_dirty_for_io(pages[i]);
1732                         if (page_ops & PAGE_SET_WRITEBACK)
1733                                 set_page_writeback(pages[i]);
1734                         if (page_ops & PAGE_SET_ERROR)
1735                                 SetPageError(pages[i]);
1736                         if (page_ops & PAGE_END_WRITEBACK)
1737                                 end_page_writeback(pages[i]);
1738                         if (page_ops & PAGE_UNLOCK)
1739                                 unlock_page(pages[i]);
1740                         if (page_ops & PAGE_LOCK) {
1741                                 lock_page(pages[i]);
1742                                 if (!PageDirty(pages[i]) ||
1743                                     pages[i]->mapping != mapping) {
1744                                         unlock_page(pages[i]);
1745                                         put_page(pages[i]);
1746                                         err = -EAGAIN;
1747                                         goto out;
1748                                 }
1749                         }
1750                         put_page(pages[i]);
1751                         pages_locked++;
1752                 }
1753                 nr_pages -= ret;
1754                 index += ret;
1755                 cond_resched();
1756         }
1757 out:
1758         if (err && index_ret)
1759                 *index_ret = start_index + pages_locked - 1;
1760         return err;
1761 }
1762
1763 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1764                                  u64 delalloc_end, struct page *locked_page,
1765                                  unsigned clear_bits,
1766                                  unsigned long page_ops)
1767 {
1768         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1769                          NULL, GFP_NOFS);
1770
1771         __process_pages_contig(inode->i_mapping, locked_page,
1772                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1773                                page_ops, NULL);
1774 }
1775
1776 /*
1777  * count the number of bytes in the tree that have a given bit(s)
1778  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1779  * cached.  The total number found is returned.
1780  */
1781 u64 count_range_bits(struct extent_io_tree *tree,
1782                      u64 *start, u64 search_end, u64 max_bytes,
1783                      unsigned bits, int contig)
1784 {
1785         struct rb_node *node;
1786         struct extent_state *state;
1787         u64 cur_start = *start;
1788         u64 total_bytes = 0;
1789         u64 last = 0;
1790         int found = 0;
1791
1792         if (WARN_ON(search_end <= cur_start))
1793                 return 0;
1794
1795         spin_lock(&tree->lock);
1796         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1797                 total_bytes = tree->dirty_bytes;
1798                 goto out;
1799         }
1800         /*
1801          * this search will find all the extents that end after
1802          * our range starts.
1803          */
1804         node = tree_search(tree, cur_start);
1805         if (!node)
1806                 goto out;
1807
1808         while (1) {
1809                 state = rb_entry(node, struct extent_state, rb_node);
1810                 if (state->start > search_end)
1811                         break;
1812                 if (contig && found && state->start > last + 1)
1813                         break;
1814                 if (state->end >= cur_start && (state->state & bits) == bits) {
1815                         total_bytes += min(search_end, state->end) + 1 -
1816                                        max(cur_start, state->start);
1817                         if (total_bytes >= max_bytes)
1818                                 break;
1819                         if (!found) {
1820                                 *start = max(cur_start, state->start);
1821                                 found = 1;
1822                         }
1823                         last = state->end;
1824                 } else if (contig && found) {
1825                         break;
1826                 }
1827                 node = rb_next(node);
1828                 if (!node)
1829                         break;
1830         }
1831 out:
1832         spin_unlock(&tree->lock);
1833         return total_bytes;
1834 }
1835
1836 /*
1837  * set the private field for a given byte offset in the tree.  If there isn't
1838  * an extent_state there already, this does nothing.
1839  */
1840 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1841                 struct io_failure_record *failrec)
1842 {
1843         struct rb_node *node;
1844         struct extent_state *state;
1845         int ret = 0;
1846
1847         spin_lock(&tree->lock);
1848         /*
1849          * this search will find all the extents that end after
1850          * our range starts.
1851          */
1852         node = tree_search(tree, start);
1853         if (!node) {
1854                 ret = -ENOENT;
1855                 goto out;
1856         }
1857         state = rb_entry(node, struct extent_state, rb_node);
1858         if (state->start != start) {
1859                 ret = -ENOENT;
1860                 goto out;
1861         }
1862         state->failrec = failrec;
1863 out:
1864         spin_unlock(&tree->lock);
1865         return ret;
1866 }
1867
1868 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1869                 struct io_failure_record **failrec)
1870 {
1871         struct rb_node *node;
1872         struct extent_state *state;
1873         int ret = 0;
1874
1875         spin_lock(&tree->lock);
1876         /*
1877          * this search will find all the extents that end after
1878          * our range starts.
1879          */
1880         node = tree_search(tree, start);
1881         if (!node) {
1882                 ret = -ENOENT;
1883                 goto out;
1884         }
1885         state = rb_entry(node, struct extent_state, rb_node);
1886         if (state->start != start) {
1887                 ret = -ENOENT;
1888                 goto out;
1889         }
1890         *failrec = state->failrec;
1891 out:
1892         spin_unlock(&tree->lock);
1893         return ret;
1894 }
1895
1896 /*
1897  * searches a range in the state tree for a given mask.
1898  * If 'filled' == 1, this returns 1 only if every extent in the tree
1899  * has the bits set.  Otherwise, 1 is returned if any bit in the
1900  * range is found set.
1901  */
1902 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1903                    unsigned bits, int filled, struct extent_state *cached)
1904 {
1905         struct extent_state *state = NULL;
1906         struct rb_node *node;
1907         int bitset = 0;
1908
1909         spin_lock(&tree->lock);
1910         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1911             cached->end > start)
1912                 node = &cached->rb_node;
1913         else
1914                 node = tree_search(tree, start);
1915         while (node && start <= end) {
1916                 state = rb_entry(node, struct extent_state, rb_node);
1917
1918                 if (filled && state->start > start) {
1919                         bitset = 0;
1920                         break;
1921                 }
1922
1923                 if (state->start > end)
1924                         break;
1925
1926                 if (state->state & bits) {
1927                         bitset = 1;
1928                         if (!filled)
1929                                 break;
1930                 } else if (filled) {
1931                         bitset = 0;
1932                         break;
1933                 }
1934
1935                 if (state->end == (u64)-1)
1936                         break;
1937
1938                 start = state->end + 1;
1939                 if (start > end)
1940                         break;
1941                 node = rb_next(node);
1942                 if (!node) {
1943                         if (filled)
1944                                 bitset = 0;
1945                         break;
1946                 }
1947         }
1948         spin_unlock(&tree->lock);
1949         return bitset;
1950 }
1951
1952 /*
1953  * helper function to set a given page up to date if all the
1954  * extents in the tree for that page are up to date
1955  */
1956 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1957 {
1958         u64 start = page_offset(page);
1959         u64 end = start + PAGE_SIZE - 1;
1960         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1961                 SetPageUptodate(page);
1962 }
1963
1964 int free_io_failure(struct btrfs_inode *inode, struct io_failure_record *rec)
1965 {
1966         int ret;
1967         int err = 0;
1968         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
1969
1970         set_state_failrec(failure_tree, rec->start, NULL);
1971         ret = clear_extent_bits(failure_tree, rec->start,
1972                                 rec->start + rec->len - 1,
1973                                 EXTENT_LOCKED | EXTENT_DIRTY);
1974         if (ret)
1975                 err = ret;
1976
1977         ret = clear_extent_bits(&inode->io_tree, rec->start,
1978                                 rec->start + rec->len - 1,
1979                                 EXTENT_DAMAGED);
1980         if (ret && !err)
1981                 err = ret;
1982
1983         kfree(rec);
1984         return err;
1985 }
1986
1987 /*
1988  * this bypasses the standard btrfs submit functions deliberately, as
1989  * the standard behavior is to write all copies in a raid setup. here we only
1990  * want to write the one bad copy. so we do the mapping for ourselves and issue
1991  * submit_bio directly.
1992  * to avoid any synchronization issues, wait for the data after writing, which
1993  * actually prevents the read that triggered the error from finishing.
1994  * currently, there can be no more than two copies of every data bit. thus,
1995  * exactly one rewrite is required.
1996  */
1997 int repair_io_failure(struct btrfs_inode *inode, u64 start, u64 length,
1998                 u64 logical, struct page *page,
1999                 unsigned int pg_offset, int mirror_num)
2000 {
2001         struct btrfs_fs_info *fs_info = inode->root->fs_info;
2002         struct bio *bio;
2003         struct btrfs_device *dev;
2004         u64 map_length = 0;
2005         u64 sector;
2006         struct btrfs_bio *bbio = NULL;
2007         int ret;
2008
2009         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2010         BUG_ON(!mirror_num);
2011
2012         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2013         if (!bio)
2014                 return -EIO;
2015         bio->bi_iter.bi_size = 0;
2016         map_length = length;
2017
2018         /*
2019          * Avoid races with device replace and make sure our bbio has devices
2020          * associated to its stripes that don't go away while we are doing the
2021          * read repair operation.
2022          */
2023         btrfs_bio_counter_inc_blocked(fs_info);
2024         if (btrfs_is_parity_mirror(fs_info, logical, length, mirror_num)) {
2025                 /*
2026                  * Note that we don't use BTRFS_MAP_WRITE because it's supposed
2027                  * to update all raid stripes, but here we just want to correct
2028                  * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
2029                  * stripe's dev and sector.
2030                  */
2031                 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
2032                                       &map_length, &bbio, 0);
2033                 if (ret) {
2034                         btrfs_bio_counter_dec(fs_info);
2035                         bio_put(bio);
2036                         return -EIO;
2037                 }
2038                 ASSERT(bbio->mirror_num == 1);
2039         } else {
2040                 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2041                                       &map_length, &bbio, mirror_num);
2042                 if (ret) {
2043                         btrfs_bio_counter_dec(fs_info);
2044                         bio_put(bio);
2045                         return -EIO;
2046                 }
2047                 BUG_ON(mirror_num != bbio->mirror_num);
2048         }
2049
2050         sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2051         bio->bi_iter.bi_sector = sector;
2052         dev = bbio->stripes[bbio->mirror_num - 1].dev;
2053         btrfs_put_bbio(bbio);
2054         if (!dev || !dev->bdev || !dev->writeable) {
2055                 btrfs_bio_counter_dec(fs_info);
2056                 bio_put(bio);
2057                 return -EIO;
2058         }
2059         bio->bi_bdev = dev->bdev;
2060         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2061         bio_add_page(bio, page, length, pg_offset);
2062
2063         if (btrfsic_submit_bio_wait(bio)) {
2064                 /* try to remap that extent elsewhere? */
2065                 btrfs_bio_counter_dec(fs_info);
2066                 bio_put(bio);
2067                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2068                 return -EIO;
2069         }
2070
2071         btrfs_info_rl_in_rcu(fs_info,
2072                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2073                                   btrfs_ino(inode), start,
2074                                   rcu_str_deref(dev->name), sector);
2075         btrfs_bio_counter_dec(fs_info);
2076         bio_put(bio);
2077         return 0;
2078 }
2079
2080 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2081                          struct extent_buffer *eb, int mirror_num)
2082 {
2083         u64 start = eb->start;
2084         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2085         int ret = 0;
2086
2087         if (fs_info->sb->s_flags & MS_RDONLY)
2088                 return -EROFS;
2089
2090         for (i = 0; i < num_pages; i++) {
2091                 struct page *p = eb->pages[i];
2092
2093                 ret = repair_io_failure(BTRFS_I(fs_info->btree_inode), start,
2094                                         PAGE_SIZE, start, p,
2095                                         start - page_offset(p), mirror_num);
2096                 if (ret)
2097                         break;
2098                 start += PAGE_SIZE;
2099         }
2100
2101         return ret;
2102 }
2103
2104 /*
2105  * each time an IO finishes, we do a fast check in the IO failure tree
2106  * to see if we need to process or clean up an io_failure_record
2107  */
2108 int clean_io_failure(struct btrfs_inode *inode, u64 start, struct page *page,
2109                      unsigned int pg_offset)
2110 {
2111         u64 private;
2112         struct io_failure_record *failrec;
2113         struct btrfs_fs_info *fs_info = inode->root->fs_info;
2114         struct extent_state *state;
2115         int num_copies;
2116         int ret;
2117
2118         private = 0;
2119         ret = count_range_bits(&inode->io_failure_tree, &private,
2120                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2121         if (!ret)
2122                 return 0;
2123
2124         ret = get_state_failrec(&inode->io_failure_tree, start,
2125                         &failrec);
2126         if (ret)
2127                 return 0;
2128
2129         BUG_ON(!failrec->this_mirror);
2130
2131         if (failrec->in_validation) {
2132                 /* there was no real error, just free the record */
2133                 btrfs_debug(fs_info,
2134                         "clean_io_failure: freeing dummy error at %llu",
2135                         failrec->start);
2136                 goto out;
2137         }
2138         if (fs_info->sb->s_flags & MS_RDONLY)
2139                 goto out;
2140
2141         spin_lock(&inode->io_tree.lock);
2142         state = find_first_extent_bit_state(&inode->io_tree,
2143                                             failrec->start,
2144                                             EXTENT_LOCKED);
2145         spin_unlock(&inode->io_tree.lock);
2146
2147         if (state && state->start <= failrec->start &&
2148             state->end >= failrec->start + failrec->len - 1) {
2149                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2150                                               failrec->len);
2151                 if (num_copies > 1)  {
2152                         repair_io_failure(inode, start, failrec->len,
2153                                           failrec->logical, page,
2154                                           pg_offset, failrec->failed_mirror);
2155                 }
2156         }
2157
2158 out:
2159         free_io_failure(inode, failrec);
2160
2161         return 0;
2162 }
2163
2164 /*
2165  * Can be called when
2166  * - hold extent lock
2167  * - under ordered extent
2168  * - the inode is freeing
2169  */
2170 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2171 {
2172         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2173         struct io_failure_record *failrec;
2174         struct extent_state *state, *next;
2175
2176         if (RB_EMPTY_ROOT(&failure_tree->state))
2177                 return;
2178
2179         spin_lock(&failure_tree->lock);
2180         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2181         while (state) {
2182                 if (state->start > end)
2183                         break;
2184
2185                 ASSERT(state->end <= end);
2186
2187                 next = next_state(state);
2188
2189                 failrec = state->failrec;
2190                 free_extent_state(state);
2191                 kfree(failrec);
2192
2193                 state = next;
2194         }
2195         spin_unlock(&failure_tree->lock);
2196 }
2197
2198 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2199                 struct io_failure_record **failrec_ret)
2200 {
2201         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2202         struct io_failure_record *failrec;
2203         struct extent_map *em;
2204         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2205         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2206         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2207         int ret;
2208         u64 logical;
2209
2210         ret = get_state_failrec(failure_tree, start, &failrec);
2211         if (ret) {
2212                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2213                 if (!failrec)
2214                         return -ENOMEM;
2215
2216                 failrec->start = start;
2217                 failrec->len = end - start + 1;
2218                 failrec->this_mirror = 0;
2219                 failrec->bio_flags = 0;
2220                 failrec->in_validation = 0;
2221
2222                 read_lock(&em_tree->lock);
2223                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2224                 if (!em) {
2225                         read_unlock(&em_tree->lock);
2226                         kfree(failrec);
2227                         return -EIO;
2228                 }
2229
2230                 if (em->start > start || em->start + em->len <= start) {
2231                         free_extent_map(em);
2232                         em = NULL;
2233                 }
2234                 read_unlock(&em_tree->lock);
2235                 if (!em) {
2236                         kfree(failrec);
2237                         return -EIO;
2238                 }
2239
2240                 logical = start - em->start;
2241                 logical = em->block_start + logical;
2242                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2243                         logical = em->block_start;
2244                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2245                         extent_set_compress_type(&failrec->bio_flags,
2246                                                  em->compress_type);
2247                 }
2248
2249                 btrfs_debug(fs_info,
2250                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2251                         logical, start, failrec->len);
2252
2253                 failrec->logical = logical;
2254                 free_extent_map(em);
2255
2256                 /* set the bits in the private failure tree */
2257                 ret = set_extent_bits(failure_tree, start, end,
2258                                         EXTENT_LOCKED | EXTENT_DIRTY);
2259                 if (ret >= 0)
2260                         ret = set_state_failrec(failure_tree, start, failrec);
2261                 /* set the bits in the inode's tree */
2262                 if (ret >= 0)
2263                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2264                 if (ret < 0) {
2265                         kfree(failrec);
2266                         return ret;
2267                 }
2268         } else {
2269                 btrfs_debug(fs_info,
2270                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2271                         failrec->logical, failrec->start, failrec->len,
2272                         failrec->in_validation);
2273                 /*
2274                  * when data can be on disk more than twice, add to failrec here
2275                  * (e.g. with a list for failed_mirror) to make
2276                  * clean_io_failure() clean all those errors at once.
2277                  */
2278         }
2279
2280         *failrec_ret = failrec;
2281
2282         return 0;
2283 }
2284
2285 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2286                            struct io_failure_record *failrec, int failed_mirror)
2287 {
2288         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2289         int num_copies;
2290
2291         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2292         if (num_copies == 1) {
2293                 /*
2294                  * we only have a single copy of the data, so don't bother with
2295                  * all the retry and error correction code that follows. no
2296                  * matter what the error is, it is very likely to persist.
2297                  */
2298                 btrfs_debug(fs_info,
2299                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2300                         num_copies, failrec->this_mirror, failed_mirror);
2301                 return 0;
2302         }
2303
2304         /*
2305          * there are two premises:
2306          *      a) deliver good data to the caller
2307          *      b) correct the bad sectors on disk
2308          */
2309         if (failed_bio->bi_vcnt > 1) {
2310                 /*
2311                  * to fulfill b), we need to know the exact failing sectors, as
2312                  * we don't want to rewrite any more than the failed ones. thus,
2313                  * we need separate read requests for the failed bio
2314                  *
2315                  * if the following BUG_ON triggers, our validation request got
2316                  * merged. we need separate requests for our algorithm to work.
2317                  */
2318                 BUG_ON(failrec->in_validation);
2319                 failrec->in_validation = 1;
2320                 failrec->this_mirror = failed_mirror;
2321         } else {
2322                 /*
2323                  * we're ready to fulfill a) and b) alongside. get a good copy
2324                  * of the failed sector and if we succeed, we have setup
2325                  * everything for repair_io_failure to do the rest for us.
2326                  */
2327                 if (failrec->in_validation) {
2328                         BUG_ON(failrec->this_mirror != failed_mirror);
2329                         failrec->in_validation = 0;
2330                         failrec->this_mirror = 0;
2331                 }
2332                 failrec->failed_mirror = failed_mirror;
2333                 failrec->this_mirror++;
2334                 if (failrec->this_mirror == failed_mirror)
2335                         failrec->this_mirror++;
2336         }
2337
2338         if (failrec->this_mirror > num_copies) {
2339                 btrfs_debug(fs_info,
2340                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2341                         num_copies, failrec->this_mirror, failed_mirror);
2342                 return 0;
2343         }
2344
2345         return 1;
2346 }
2347
2348
2349 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2350                                     struct io_failure_record *failrec,
2351                                     struct page *page, int pg_offset, int icsum,
2352                                     bio_end_io_t *endio_func, void *data)
2353 {
2354         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2355         struct bio *bio;
2356         struct btrfs_io_bio *btrfs_failed_bio;
2357         struct btrfs_io_bio *btrfs_bio;
2358
2359         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2360         if (!bio)
2361                 return NULL;
2362
2363         bio->bi_end_io = endio_func;
2364         bio->bi_iter.bi_sector = failrec->logical >> 9;
2365         bio->bi_bdev = fs_info->fs_devices->latest_bdev;
2366         bio->bi_iter.bi_size = 0;
2367         bio->bi_private = data;
2368
2369         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2370         if (btrfs_failed_bio->csum) {
2371                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2372
2373                 btrfs_bio = btrfs_io_bio(bio);
2374                 btrfs_bio->csum = btrfs_bio->csum_inline;
2375                 icsum *= csum_size;
2376                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2377                        csum_size);
2378         }
2379
2380         bio_add_page(bio, page, failrec->len, pg_offset);
2381
2382         return bio;
2383 }
2384
2385 /*
2386  * this is a generic handler for readpage errors (default
2387  * readpage_io_failed_hook). if other copies exist, read those and write back
2388  * good data to the failed position. does not investigate in remapping the
2389  * failed extent elsewhere, hoping the device will be smart enough to do this as
2390  * needed
2391  */
2392
2393 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2394                               struct page *page, u64 start, u64 end,
2395                               int failed_mirror)
2396 {
2397         struct io_failure_record *failrec;
2398         struct inode *inode = page->mapping->host;
2399         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2400         struct bio *bio;
2401         int read_mode = 0;
2402         int ret;
2403
2404         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2405
2406         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2407         if (ret)
2408                 return ret;
2409
2410         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2411         if (!ret) {
2412                 free_io_failure(BTRFS_I(inode), failrec);
2413                 return -EIO;
2414         }
2415
2416         if (failed_bio->bi_vcnt > 1)
2417                 read_mode |= REQ_FAILFAST_DEV;
2418
2419         phy_offset >>= inode->i_sb->s_blocksize_bits;
2420         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2421                                       start - page_offset(page),
2422                                       (int)phy_offset, failed_bio->bi_end_io,
2423                                       NULL);
2424         if (!bio) {
2425                 free_io_failure(BTRFS_I(inode), failrec);
2426                 return -EIO;
2427         }
2428         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2429
2430         btrfs_debug(btrfs_sb(inode->i_sb),
2431                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2432                 read_mode, failrec->this_mirror, failrec->in_validation);
2433
2434         ret = tree->ops->submit_bio_hook(inode, bio, failrec->this_mirror,
2435                                          failrec->bio_flags, 0);
2436         if (ret) {
2437                 free_io_failure(BTRFS_I(inode), failrec);
2438                 bio_put(bio);
2439         }
2440
2441         return ret;
2442 }
2443
2444 /* lots and lots of room for performance fixes in the end_bio funcs */
2445
2446 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2447 {
2448         int uptodate = (err == 0);
2449         struct extent_io_tree *tree;
2450         int ret = 0;
2451
2452         tree = &BTRFS_I(page->mapping->host)->io_tree;
2453
2454         if (tree->ops && tree->ops->writepage_end_io_hook)
2455                 tree->ops->writepage_end_io_hook(page, start, end, NULL,
2456                                 uptodate);
2457
2458         if (!uptodate) {
2459                 ClearPageUptodate(page);
2460                 SetPageError(page);
2461                 ret = ret < 0 ? ret : -EIO;
2462                 mapping_set_error(page->mapping, ret);
2463         }
2464 }
2465
2466 /*
2467  * after a writepage IO is done, we need to:
2468  * clear the uptodate bits on error
2469  * clear the writeback bits in the extent tree for this IO
2470  * end_page_writeback if the page has no more pending IO
2471  *
2472  * Scheduling is not allowed, so the extent state tree is expected
2473  * to have one and only one object corresponding to this IO.
2474  */
2475 static void end_bio_extent_writepage(struct bio *bio)
2476 {
2477         struct bio_vec *bvec;
2478         u64 start;
2479         u64 end;
2480         int i;
2481
2482         bio_for_each_segment_all(bvec, bio, i) {
2483                 struct page *page = bvec->bv_page;
2484                 struct inode *inode = page->mapping->host;
2485                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2486
2487                 /* We always issue full-page reads, but if some block
2488                  * in a page fails to read, blk_update_request() will
2489                  * advance bv_offset and adjust bv_len to compensate.
2490                  * Print a warning for nonzero offsets, and an error
2491                  * if they don't add up to a full page.  */
2492                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2493                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2494                                 btrfs_err(fs_info,
2495                                    "partial page write in btrfs with offset %u and length %u",
2496                                         bvec->bv_offset, bvec->bv_len);
2497                         else
2498                                 btrfs_info(fs_info,
2499                                    "incomplete page write in btrfs with offset %u and length %u",
2500                                         bvec->bv_offset, bvec->bv_len);
2501                 }
2502
2503                 start = page_offset(page);
2504                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2505
2506                 end_extent_writepage(page, bio->bi_error, start, end);
2507                 end_page_writeback(page);
2508         }
2509
2510         bio_put(bio);
2511 }
2512
2513 static void
2514 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2515                               int uptodate)
2516 {
2517         struct extent_state *cached = NULL;
2518         u64 end = start + len - 1;
2519
2520         if (uptodate && tree->track_uptodate)
2521                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2522         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2523 }
2524
2525 /*
2526  * after a readpage IO is done, we need to:
2527  * clear the uptodate bits on error
2528  * set the uptodate bits if things worked
2529  * set the page up to date if all extents in the tree are uptodate
2530  * clear the lock bit in the extent tree
2531  * unlock the page if there are no other extents locked for it
2532  *
2533  * Scheduling is not allowed, so the extent state tree is expected
2534  * to have one and only one object corresponding to this IO.
2535  */
2536 static void end_bio_extent_readpage(struct bio *bio)
2537 {
2538         struct bio_vec *bvec;
2539         int uptodate = !bio->bi_error;
2540         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2541         struct extent_io_tree *tree;
2542         u64 offset = 0;
2543         u64 start;
2544         u64 end;
2545         u64 len;
2546         u64 extent_start = 0;
2547         u64 extent_len = 0;
2548         int mirror;
2549         int ret;
2550         int i;
2551
2552         bio_for_each_segment_all(bvec, bio, i) {
2553                 struct page *page = bvec->bv_page;
2554                 struct inode *inode = page->mapping->host;
2555                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2556
2557                 btrfs_debug(fs_info,
2558                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2559                         (u64)bio->bi_iter.bi_sector, bio->bi_error,
2560                         io_bio->mirror_num);
2561                 tree = &BTRFS_I(inode)->io_tree;
2562
2563                 /* We always issue full-page reads, but if some block
2564                  * in a page fails to read, blk_update_request() will
2565                  * advance bv_offset and adjust bv_len to compensate.
2566                  * Print a warning for nonzero offsets, and an error
2567                  * if they don't add up to a full page.  */
2568                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2569                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2570                                 btrfs_err(fs_info,
2571                                         "partial page read in btrfs with offset %u and length %u",
2572                                         bvec->bv_offset, bvec->bv_len);
2573                         else
2574                                 btrfs_info(fs_info,
2575                                         "incomplete page read in btrfs with offset %u and length %u",
2576                                         bvec->bv_offset, bvec->bv_len);
2577                 }
2578
2579                 start = page_offset(page);
2580                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2581                 len = bvec->bv_len;
2582
2583                 mirror = io_bio->mirror_num;
2584                 if (likely(uptodate && tree->ops)) {
2585                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2586                                                               page, start, end,
2587                                                               mirror);
2588                         if (ret)
2589                                 uptodate = 0;
2590                         else
2591                                 clean_io_failure(BTRFS_I(inode), start,
2592                                                 page, 0);
2593                 }
2594
2595                 if (likely(uptodate))
2596                         goto readpage_ok;
2597
2598                 if (tree->ops) {
2599                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2600                         if (ret == -EAGAIN) {
2601                                 /*
2602                                  * Data inode's readpage_io_failed_hook() always
2603                                  * returns -EAGAIN.
2604                                  *
2605                                  * The generic bio_readpage_error handles errors
2606                                  * the following way: If possible, new read
2607                                  * requests are created and submitted and will
2608                                  * end up in end_bio_extent_readpage as well (if
2609                                  * we're lucky, not in the !uptodate case). In
2610                                  * that case it returns 0 and we just go on with
2611                                  * the next page in our bio. If it can't handle
2612                                  * the error it will return -EIO and we remain
2613                                  * responsible for that page.
2614                                  */
2615                                 ret = bio_readpage_error(bio, offset, page,
2616                                                          start, end, mirror);
2617                                 if (ret == 0) {
2618                                         uptodate = !bio->bi_error;
2619                                         offset += len;
2620                                         continue;
2621                                 }
2622                         }
2623
2624                         /*
2625                          * metadata's readpage_io_failed_hook() always returns
2626                          * -EIO and fixes nothing.  -EIO is also returned if
2627                          * data inode error could not be fixed.
2628                          */
2629                         ASSERT(ret == -EIO);
2630                 }
2631 readpage_ok:
2632                 if (likely(uptodate)) {
2633                         loff_t i_size = i_size_read(inode);
2634                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2635                         unsigned off;
2636
2637                         /* Zero out the end if this page straddles i_size */
2638                         off = i_size & (PAGE_SIZE-1);
2639                         if (page->index == end_index && off)
2640                                 zero_user_segment(page, off, PAGE_SIZE);
2641                         SetPageUptodate(page);
2642                 } else {
2643                         ClearPageUptodate(page);
2644                         SetPageError(page);
2645                 }
2646                 unlock_page(page);
2647                 offset += len;
2648
2649                 if (unlikely(!uptodate)) {
2650                         if (extent_len) {
2651                                 endio_readpage_release_extent(tree,
2652                                                               extent_start,
2653                                                               extent_len, 1);
2654                                 extent_start = 0;
2655                                 extent_len = 0;
2656                         }
2657                         endio_readpage_release_extent(tree, start,
2658                                                       end - start + 1, 0);
2659                 } else if (!extent_len) {
2660                         extent_start = start;
2661                         extent_len = end + 1 - start;
2662                 } else if (extent_start + extent_len == start) {
2663                         extent_len += end + 1 - start;
2664                 } else {
2665                         endio_readpage_release_extent(tree, extent_start,
2666                                                       extent_len, uptodate);
2667                         extent_start = start;
2668                         extent_len = end + 1 - start;
2669                 }
2670         }
2671
2672         if (extent_len)
2673                 endio_readpage_release_extent(tree, extent_start, extent_len,
2674                                               uptodate);
2675         if (io_bio->end_io)
2676                 io_bio->end_io(io_bio, bio->bi_error);
2677         bio_put(bio);
2678 }
2679
2680 /*
2681  * this allocates from the btrfs_bioset.  We're returning a bio right now
2682  * but you can call btrfs_io_bio for the appropriate container_of magic
2683  */
2684 struct bio *
2685 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2686                 gfp_t gfp_flags)
2687 {
2688         struct btrfs_io_bio *btrfs_bio;
2689         struct bio *bio;
2690
2691         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2692
2693         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2694                 while (!bio && (nr_vecs /= 2)) {
2695                         bio = bio_alloc_bioset(gfp_flags,
2696                                                nr_vecs, btrfs_bioset);
2697                 }
2698         }
2699
2700         if (bio) {
2701                 bio->bi_bdev = bdev;
2702                 bio->bi_iter.bi_sector = first_sector;
2703                 btrfs_bio = btrfs_io_bio(bio);
2704                 btrfs_bio->csum = NULL;
2705                 btrfs_bio->csum_allocated = NULL;
2706                 btrfs_bio->end_io = NULL;
2707         }
2708         return bio;
2709 }
2710
2711 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2712 {
2713         struct btrfs_io_bio *btrfs_bio;
2714         struct bio *new;
2715
2716         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2717         if (new) {
2718                 btrfs_bio = btrfs_io_bio(new);
2719                 btrfs_bio->csum = NULL;
2720                 btrfs_bio->csum_allocated = NULL;
2721                 btrfs_bio->end_io = NULL;
2722         }
2723         return new;
2724 }
2725
2726 /* this also allocates from the btrfs_bioset */
2727 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2728 {
2729         struct btrfs_io_bio *btrfs_bio;
2730         struct bio *bio;
2731
2732         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2733         if (bio) {
2734                 btrfs_bio = btrfs_io_bio(bio);
2735                 btrfs_bio->csum = NULL;
2736                 btrfs_bio->csum_allocated = NULL;
2737                 btrfs_bio->end_io = NULL;
2738         }
2739         return bio;
2740 }
2741
2742
2743 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2744                                        unsigned long bio_flags)
2745 {
2746         int ret = 0;
2747         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2748         struct page *page = bvec->bv_page;
2749         struct extent_io_tree *tree = bio->bi_private;
2750         u64 start;
2751
2752         start = page_offset(page) + bvec->bv_offset;
2753
2754         bio->bi_private = NULL;
2755         bio_get(bio);
2756
2757         if (tree->ops)
2758                 ret = tree->ops->submit_bio_hook(page->mapping->host, bio,
2759                                            mirror_num, bio_flags, start);
2760         else
2761                 btrfsic_submit_bio(bio);
2762
2763         bio_put(bio);
2764         return ret;
2765 }
2766
2767 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2768                      unsigned long offset, size_t size, struct bio *bio,
2769                      unsigned long bio_flags)
2770 {
2771         int ret = 0;
2772         if (tree->ops)
2773                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2774                                                 bio_flags);
2775         return ret;
2776
2777 }
2778
2779 static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree,
2780                               struct writeback_control *wbc,
2781                               struct page *page, sector_t sector,
2782                               size_t size, unsigned long offset,
2783                               struct block_device *bdev,
2784                               struct bio **bio_ret,
2785                               bio_end_io_t end_io_func,
2786                               int mirror_num,
2787                               unsigned long prev_bio_flags,
2788                               unsigned long bio_flags,
2789                               bool force_bio_submit)
2790 {
2791         int ret = 0;
2792         struct bio *bio;
2793         int contig = 0;
2794         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2795         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2796
2797         if (bio_ret && *bio_ret) {
2798                 bio = *bio_ret;
2799                 if (old_compressed)
2800                         contig = bio->bi_iter.bi_sector == sector;
2801                 else
2802                         contig = bio_end_sector(bio) == sector;
2803
2804                 if (prev_bio_flags != bio_flags || !contig ||
2805                     force_bio_submit ||
2806                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2807                     bio_add_page(bio, page, page_size, offset) < page_size) {
2808                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2809                         if (ret < 0) {
2810                                 *bio_ret = NULL;
2811                                 return ret;
2812                         }
2813                         bio = NULL;
2814                 } else {
2815                         if (wbc)
2816                                 wbc_account_io(wbc, page, page_size);
2817                         return 0;
2818                 }
2819         }
2820
2821         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2822                         GFP_NOFS | __GFP_HIGH);
2823         if (!bio)
2824                 return -ENOMEM;
2825
2826         bio_add_page(bio, page, page_size, offset);
2827         bio->bi_end_io = end_io_func;
2828         bio->bi_private = tree;
2829         bio_set_op_attrs(bio, op, op_flags);
2830         if (wbc) {
2831                 wbc_init_bio(wbc, bio);
2832                 wbc_account_io(wbc, page, page_size);
2833         }
2834
2835         if (bio_ret)
2836                 *bio_ret = bio;
2837         else
2838                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2839
2840         return ret;
2841 }
2842
2843 static void attach_extent_buffer_page(struct extent_buffer *eb,
2844                                       struct page *page)
2845 {
2846         if (!PagePrivate(page)) {
2847                 SetPagePrivate(page);
2848                 get_page(page);
2849                 set_page_private(page, (unsigned long)eb);
2850         } else {
2851                 WARN_ON(page->private != (unsigned long)eb);
2852         }
2853 }
2854
2855 void set_page_extent_mapped(struct page *page)
2856 {
2857         if (!PagePrivate(page)) {
2858                 SetPagePrivate(page);
2859                 get_page(page);
2860                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2861         }
2862 }
2863
2864 static struct extent_map *
2865 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2866                  u64 start, u64 len, get_extent_t *get_extent,
2867                  struct extent_map **em_cached)
2868 {
2869         struct extent_map *em;
2870
2871         if (em_cached && *em_cached) {
2872                 em = *em_cached;
2873                 if (extent_map_in_tree(em) && start >= em->start &&
2874                     start < extent_map_end(em)) {
2875                         refcount_inc(&em->refs);
2876                         return em;
2877                 }
2878
2879                 free_extent_map(em);
2880                 *em_cached = NULL;
2881         }
2882
2883         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2884         if (em_cached && !IS_ERR_OR_NULL(em)) {
2885                 BUG_ON(*em_cached);
2886                 refcount_inc(&em->refs);
2887                 *em_cached = em;
2888         }
2889         return em;
2890 }
2891 /*
2892  * basic readpage implementation.  Locked extent state structs are inserted
2893  * into the tree that are removed when the IO is done (by the end_io
2894  * handlers)
2895  * XXX JDM: This needs looking at to ensure proper page locking
2896  * return 0 on success, otherwise return error
2897  */
2898 static int __do_readpage(struct extent_io_tree *tree,
2899                          struct page *page,
2900                          get_extent_t *get_extent,
2901                          struct extent_map **em_cached,
2902                          struct bio **bio, int mirror_num,
2903                          unsigned long *bio_flags, int read_flags,
2904                          u64 *prev_em_start)
2905 {
2906         struct inode *inode = page->mapping->host;
2907         u64 start = page_offset(page);
2908         u64 page_end = start + PAGE_SIZE - 1;
2909         u64 end;
2910         u64 cur = start;
2911         u64 extent_offset;
2912         u64 last_byte = i_size_read(inode);
2913         u64 block_start;
2914         u64 cur_end;
2915         sector_t sector;
2916         struct extent_map *em;
2917         struct block_device *bdev;
2918         int ret = 0;
2919         int nr = 0;
2920         size_t pg_offset = 0;
2921         size_t iosize;
2922         size_t disk_io_size;
2923         size_t blocksize = inode->i_sb->s_blocksize;
2924         unsigned long this_bio_flag = 0;
2925
2926         set_page_extent_mapped(page);
2927
2928         end = page_end;
2929         if (!PageUptodate(page)) {
2930                 if (cleancache_get_page(page) == 0) {
2931                         BUG_ON(blocksize != PAGE_SIZE);
2932                         unlock_extent(tree, start, end);
2933                         goto out;
2934                 }
2935         }
2936
2937         if (page->index == last_byte >> PAGE_SHIFT) {
2938                 char *userpage;
2939                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2940
2941                 if (zero_offset) {
2942                         iosize = PAGE_SIZE - zero_offset;
2943                         userpage = kmap_atomic(page);
2944                         memset(userpage + zero_offset, 0, iosize);
2945                         flush_dcache_page(page);
2946                         kunmap_atomic(userpage);
2947                 }
2948         }
2949         while (cur <= end) {
2950                 bool force_bio_submit = false;
2951
2952                 if (cur >= last_byte) {
2953                         char *userpage;
2954                         struct extent_state *cached = NULL;
2955
2956                         iosize = PAGE_SIZE - pg_offset;
2957                         userpage = kmap_atomic(page);
2958                         memset(userpage + pg_offset, 0, iosize);
2959                         flush_dcache_page(page);
2960                         kunmap_atomic(userpage);
2961                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2962                                             &cached, GFP_NOFS);
2963                         unlock_extent_cached(tree, cur,
2964                                              cur + iosize - 1,
2965                                              &cached, GFP_NOFS);
2966                         break;
2967                 }
2968                 em = __get_extent_map(inode, page, pg_offset, cur,
2969                                       end - cur + 1, get_extent, em_cached);
2970                 if (IS_ERR_OR_NULL(em)) {
2971                         SetPageError(page);
2972                         unlock_extent(tree, cur, end);
2973                         break;
2974                 }
2975                 extent_offset = cur - em->start;
2976                 BUG_ON(extent_map_end(em) <= cur);
2977                 BUG_ON(end < cur);
2978
2979                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2980                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2981                         extent_set_compress_type(&this_bio_flag,
2982                                                  em->compress_type);
2983                 }
2984
2985                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2986                 cur_end = min(extent_map_end(em) - 1, end);
2987                 iosize = ALIGN(iosize, blocksize);
2988                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2989                         disk_io_size = em->block_len;
2990                         sector = em->block_start >> 9;
2991                 } else {
2992                         sector = (em->block_start + extent_offset) >> 9;
2993                         disk_io_size = iosize;
2994                 }
2995                 bdev = em->bdev;
2996                 block_start = em->block_start;
2997                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2998                         block_start = EXTENT_MAP_HOLE;
2999
3000                 /*
3001                  * If we have a file range that points to a compressed extent
3002                  * and it's followed by a consecutive file range that points to
3003                  * to the same compressed extent (possibly with a different
3004                  * offset and/or length, so it either points to the whole extent
3005                  * or only part of it), we must make sure we do not submit a
3006                  * single bio to populate the pages for the 2 ranges because
3007                  * this makes the compressed extent read zero out the pages
3008                  * belonging to the 2nd range. Imagine the following scenario:
3009                  *
3010                  *  File layout
3011                  *  [0 - 8K]                     [8K - 24K]
3012                  *    |                               |
3013                  *    |                               |
3014                  * points to extent X,         points to extent X,
3015                  * offset 4K, length of 8K     offset 0, length 16K
3016                  *
3017                  * [extent X, compressed length = 4K uncompressed length = 16K]
3018                  *
3019                  * If the bio to read the compressed extent covers both ranges,
3020                  * it will decompress extent X into the pages belonging to the
3021                  * first range and then it will stop, zeroing out the remaining
3022                  * pages that belong to the other range that points to extent X.
3023                  * So here we make sure we submit 2 bios, one for the first
3024                  * range and another one for the third range. Both will target
3025                  * the same physical extent from disk, but we can't currently
3026                  * make the compressed bio endio callback populate the pages
3027                  * for both ranges because each compressed bio is tightly
3028                  * coupled with a single extent map, and each range can have
3029                  * an extent map with a different offset value relative to the
3030                  * uncompressed data of our extent and different lengths. This
3031                  * is a corner case so we prioritize correctness over
3032                  * non-optimal behavior (submitting 2 bios for the same extent).
3033                  */
3034                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3035                     prev_em_start && *prev_em_start != (u64)-1 &&
3036                     *prev_em_start != em->orig_start)
3037                         force_bio_submit = true;
3038
3039                 if (prev_em_start)
3040                         *prev_em_start = em->orig_start;
3041
3042                 free_extent_map(em);
3043                 em = NULL;
3044
3045                 /* we've found a hole, just zero and go on */
3046                 if (block_start == EXTENT_MAP_HOLE) {
3047                         char *userpage;
3048                         struct extent_state *cached = NULL;
3049
3050                         userpage = kmap_atomic(page);
3051                         memset(userpage + pg_offset, 0, iosize);
3052                         flush_dcache_page(page);
3053                         kunmap_atomic(userpage);
3054
3055                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3056                                             &cached, GFP_NOFS);
3057                         unlock_extent_cached(tree, cur,
3058                                              cur + iosize - 1,
3059                                              &cached, GFP_NOFS);
3060                         cur = cur + iosize;
3061                         pg_offset += iosize;
3062                         continue;
3063                 }
3064                 /* the get_extent function already copied into the page */
3065                 if (test_range_bit(tree, cur, cur_end,
3066                                    EXTENT_UPTODATE, 1, NULL)) {
3067                         check_page_uptodate(tree, page);
3068                         unlock_extent(tree, cur, cur + iosize - 1);
3069                         cur = cur + iosize;
3070                         pg_offset += iosize;
3071                         continue;
3072                 }
3073                 /* we have an inline extent but it didn't get marked up
3074                  * to date.  Error out
3075                  */
3076                 if (block_start == EXTENT_MAP_INLINE) {
3077                         SetPageError(page);
3078                         unlock_extent(tree, cur, cur + iosize - 1);
3079                         cur = cur + iosize;
3080                         pg_offset += iosize;
3081                         continue;
3082                 }
3083
3084                 ret = submit_extent_page(REQ_OP_READ, read_flags, tree, NULL,
3085                                          page, sector, disk_io_size, pg_offset,
3086                                          bdev, bio,
3087                                          end_bio_extent_readpage, mirror_num,
3088                                          *bio_flags,
3089                                          this_bio_flag,
3090                                          force_bio_submit);
3091                 if (!ret) {
3092                         nr++;
3093                         *bio_flags = this_bio_flag;
3094                 } else {
3095                         SetPageError(page);
3096                         unlock_extent(tree, cur, cur + iosize - 1);
3097                         goto out;
3098                 }
3099                 cur = cur + iosize;
3100                 pg_offset += iosize;
3101         }
3102 out:
3103         if (!nr) {
3104                 if (!PageError(page))
3105                         SetPageUptodate(page);
3106                 unlock_page(page);
3107         }
3108         return ret;
3109 }
3110
3111 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3112                                              struct page *pages[], int nr_pages,
3113                                              u64 start, u64 end,
3114                                              get_extent_t *get_extent,
3115                                              struct extent_map **em_cached,
3116                                              struct bio **bio, int mirror_num,
3117                                              unsigned long *bio_flags,
3118                                              u64 *prev_em_start)
3119 {
3120         struct inode *inode;
3121         struct btrfs_ordered_extent *ordered;
3122         int index;
3123
3124         inode = pages[0]->mapping->host;
3125         while (1) {
3126                 lock_extent(tree, start, end);
3127                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3128                                                      end - start + 1);
3129                 if (!ordered)
3130                         break;
3131                 unlock_extent(tree, start, end);
3132                 btrfs_start_ordered_extent(inode, ordered, 1);
3133                 btrfs_put_ordered_extent(ordered);
3134         }
3135
3136         for (index = 0; index < nr_pages; index++) {
3137                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3138                               mirror_num, bio_flags, 0, prev_em_start);
3139                 put_page(pages[index]);
3140         }
3141 }
3142
3143 static void __extent_readpages(struct extent_io_tree *tree,
3144                                struct page *pages[],
3145                                int nr_pages, get_extent_t *get_extent,
3146                                struct extent_map **em_cached,
3147                                struct bio **bio, int mirror_num,
3148                                unsigned long *bio_flags,
3149                                u64 *prev_em_start)
3150 {
3151         u64 start = 0;
3152         u64 end = 0;
3153         u64 page_start;
3154         int index;
3155         int first_index = 0;
3156
3157         for (index = 0; index < nr_pages; index++) {
3158                 page_start = page_offset(pages[index]);
3159                 if (!end) {
3160                         start = page_start;
3161                         end = start + PAGE_SIZE - 1;
3162                         first_index = index;
3163                 } else if (end + 1 == page_start) {
3164                         end += PAGE_SIZE;
3165                 } else {
3166                         __do_contiguous_readpages(tree, &pages[first_index],
3167                                                   index - first_index, start,
3168                                                   end, get_extent, em_cached,
3169                                                   bio, mirror_num, bio_flags,
3170                                                   prev_em_start);
3171                         start = page_start;
3172                         end = start + PAGE_SIZE - 1;
3173                         first_index = index;
3174                 }
3175         }
3176
3177         if (end)
3178                 __do_contiguous_readpages(tree, &pages[first_index],
3179                                           index - first_index, start,
3180                                           end, get_extent, em_cached, bio,
3181                                           mirror_num, bio_flags,
3182                                           prev_em_start);
3183 }
3184
3185 static int __extent_read_full_page(struct extent_io_tree *tree,
3186                                    struct page *page,
3187                                    get_extent_t *get_extent,
3188                                    struct bio **bio, int mirror_num,
3189                                    unsigned long *bio_flags, int read_flags)
3190 {
3191         struct inode *inode = page->mapping->host;
3192         struct btrfs_ordered_extent *ordered;
3193         u64 start = page_offset(page);
3194         u64 end = start + PAGE_SIZE - 1;
3195         int ret;
3196
3197         while (1) {
3198                 lock_extent(tree, start, end);
3199                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3200                                                 PAGE_SIZE);
3201                 if (!ordered)
3202                         break;
3203                 unlock_extent(tree, start, end);
3204                 btrfs_start_ordered_extent(inode, ordered, 1);
3205                 btrfs_put_ordered_extent(ordered);
3206         }
3207
3208         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3209                             bio_flags, read_flags, NULL);
3210         return ret;
3211 }
3212
3213 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3214                             get_extent_t *get_extent, int mirror_num)
3215 {
3216         struct bio *bio = NULL;
3217         unsigned long bio_flags = 0;
3218         int ret;
3219
3220         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3221                                       &bio_flags, 0);
3222         if (bio)
3223                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3224         return ret;
3225 }
3226
3227 static void update_nr_written(struct writeback_control *wbc,
3228                               unsigned long nr_written)
3229 {
3230         wbc->nr_to_write -= nr_written;
3231 }
3232
3233 /*
3234  * helper for __extent_writepage, doing all of the delayed allocation setup.
3235  *
3236  * This returns 1 if our fill_delalloc function did all the work required
3237  * to write the page (copy into inline extent).  In this case the IO has
3238  * been started and the page is already unlocked.
3239  *
3240  * This returns 0 if all went well (page still locked)
3241  * This returns < 0 if there were errors (page still locked)
3242  */
3243 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3244                               struct page *page, struct writeback_control *wbc,
3245                               struct extent_page_data *epd,
3246                               u64 delalloc_start,
3247                               unsigned long *nr_written)
3248 {
3249         struct extent_io_tree *tree = epd->tree;
3250         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3251         u64 nr_delalloc;
3252         u64 delalloc_to_write = 0;
3253         u64 delalloc_end = 0;
3254         int ret;
3255         int page_started = 0;
3256
3257         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3258                 return 0;
3259
3260         while (delalloc_end < page_end) {
3261                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3262                                                page,
3263                                                &delalloc_start,
3264                                                &delalloc_end,
3265                                                BTRFS_MAX_EXTENT_SIZE);
3266                 if (nr_delalloc == 0) {
3267                         delalloc_start = delalloc_end + 1;
3268                         continue;
3269                 }
3270                 ret = tree->ops->fill_delalloc(inode, page,
3271                                                delalloc_start,
3272                                                delalloc_end,
3273                                                &page_started,
3274                                                nr_written);
3275                 /* File system has been set read-only */
3276                 if (ret) {
3277                         SetPageError(page);
3278                         /* fill_delalloc should be return < 0 for error
3279                          * but just in case, we use > 0 here meaning the
3280                          * IO is started, so we don't want to return > 0
3281                          * unless things are going well.
3282                          */
3283                         ret = ret < 0 ? ret : -EIO;
3284                         goto done;
3285                 }
3286                 /*
3287                  * delalloc_end is already one less than the total length, so
3288                  * we don't subtract one from PAGE_SIZE
3289                  */
3290                 delalloc_to_write += (delalloc_end - delalloc_start +
3291                                       PAGE_SIZE) >> PAGE_SHIFT;
3292                 delalloc_start = delalloc_end + 1;
3293         }
3294         if (wbc->nr_to_write < delalloc_to_write) {
3295                 int thresh = 8192;
3296
3297                 if (delalloc_to_write < thresh * 2)
3298                         thresh = delalloc_to_write;
3299                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3300                                          thresh);
3301         }
3302
3303         /* did the fill delalloc function already unlock and start
3304          * the IO?
3305          */
3306         if (page_started) {
3307                 /*
3308                  * we've unlocked the page, so we can't update
3309                  * the mapping's writeback index, just update
3310                  * nr_to_write.
3311                  */
3312                 wbc->nr_to_write -= *nr_written;
3313                 return 1;
3314         }
3315
3316         ret = 0;
3317
3318 done:
3319         return ret;
3320 }
3321
3322 /*
3323  * helper for __extent_writepage.  This calls the writepage start hooks,
3324  * and does the loop to map the page into extents and bios.
3325  *
3326  * We return 1 if the IO is started and the page is unlocked,
3327  * 0 if all went well (page still locked)
3328  * < 0 if there were errors (page still locked)
3329  */
3330 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3331                                  struct page *page,
3332                                  struct writeback_control *wbc,
3333                                  struct extent_page_data *epd,
3334                                  loff_t i_size,
3335                                  unsigned long nr_written,
3336                                  int write_flags, int *nr_ret)
3337 {
3338         struct extent_io_tree *tree = epd->tree;
3339         u64 start = page_offset(page);
3340         u64 page_end = start + PAGE_SIZE - 1;
3341         u64 end;
3342         u64 cur = start;
3343         u64 extent_offset;
3344         u64 block_start;
3345         u64 iosize;
3346         sector_t sector;
3347         struct extent_map *em;
3348         struct block_device *bdev;
3349         size_t pg_offset = 0;
3350         size_t blocksize;
3351         int ret = 0;
3352         int nr = 0;
3353         bool compressed;
3354
3355         if (tree->ops && tree->ops->writepage_start_hook) {
3356                 ret = tree->ops->writepage_start_hook(page, start,
3357                                                       page_end);
3358                 if (ret) {
3359                         /* Fixup worker will requeue */
3360                         if (ret == -EBUSY)
3361                                 wbc->pages_skipped++;
3362                         else
3363                                 redirty_page_for_writepage(wbc, page);
3364
3365                         update_nr_written(wbc, nr_written);
3366                         unlock_page(page);
3367                         return 1;
3368                 }
3369         }
3370
3371         /*
3372          * we don't want to touch the inode after unlocking the page,
3373          * so we update the mapping writeback index now
3374          */
3375         update_nr_written(wbc, nr_written + 1);
3376
3377         end = page_end;
3378         if (i_size <= start) {
3379                 if (tree->ops && tree->ops->writepage_end_io_hook)
3380                         tree->ops->writepage_end_io_hook(page, start,
3381                                                          page_end, NULL, 1);
3382                 goto done;
3383         }
3384
3385         blocksize = inode->i_sb->s_blocksize;
3386
3387         while (cur <= end) {
3388                 u64 em_end;
3389
3390                 if (cur >= i_size) {
3391                         if (tree->ops && tree->ops->writepage_end_io_hook)
3392                                 tree->ops->writepage_end_io_hook(page, cur,
3393                                                          page_end, NULL, 1);
3394                         break;
3395                 }
3396                 em = epd->get_extent(BTRFS_I(inode), page, pg_offset, cur,
3397                                      end - cur + 1, 1);
3398                 if (IS_ERR_OR_NULL(em)) {
3399                         SetPageError(page);
3400                         ret = PTR_ERR_OR_ZERO(em);
3401                         break;
3402                 }
3403
3404                 extent_offset = cur - em->start;
3405                 em_end = extent_map_end(em);
3406                 BUG_ON(em_end <= cur);
3407                 BUG_ON(end < cur);
3408                 iosize = min(em_end - cur, end - cur + 1);
3409                 iosize = ALIGN(iosize, blocksize);
3410                 sector = (em->block_start + extent_offset) >> 9;
3411                 bdev = em->bdev;
3412                 block_start = em->block_start;
3413                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3414                 free_extent_map(em);
3415                 em = NULL;
3416
3417                 /*
3418                  * compressed and inline extents are written through other
3419                  * paths in the FS
3420                  */
3421                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3422                     block_start == EXTENT_MAP_INLINE) {
3423                         /*
3424                          * end_io notification does not happen here for
3425                          * compressed extents
3426                          */
3427                         if (!compressed && tree->ops &&
3428                             tree->ops->writepage_end_io_hook)
3429                                 tree->ops->writepage_end_io_hook(page, cur,
3430                                                          cur + iosize - 1,
3431                                                          NULL, 1);
3432                         else if (compressed) {
3433                                 /* we don't want to end_page_writeback on
3434                                  * a compressed extent.  this happens
3435                                  * elsewhere
3436                                  */
3437                                 nr++;
3438                         }
3439
3440                         cur += iosize;
3441                         pg_offset += iosize;
3442                         continue;
3443                 }
3444
3445                 set_range_writeback(tree, cur, cur + iosize - 1);
3446                 if (!PageWriteback(page)) {
3447                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3448                                    "page %lu not writeback, cur %llu end %llu",
3449                                page->index, cur, end);
3450                 }
3451
3452                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3453                                          page, sector, iosize, pg_offset,
3454                                          bdev, &epd->bio,
3455                                          end_bio_extent_writepage,
3456                                          0, 0, 0, false);
3457                 if (ret) {
3458                         SetPageError(page);
3459                         if (PageWriteback(page))
3460                                 end_page_writeback(page);
3461                 }
3462
3463                 cur = cur + iosize;
3464                 pg_offset += iosize;
3465                 nr++;
3466         }
3467 done:
3468         *nr_ret = nr;
3469         return ret;
3470 }
3471
3472 /*
3473  * the writepage semantics are similar to regular writepage.  extent
3474  * records are inserted to lock ranges in the tree, and as dirty areas
3475  * are found, they are marked writeback.  Then the lock bits are removed
3476  * and the end_io handler clears the writeback ranges
3477  */
3478 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3479                               void *data)
3480 {
3481         struct inode *inode = page->mapping->host;
3482         struct extent_page_data *epd = data;
3483         u64 start = page_offset(page);
3484         u64 page_end = start + PAGE_SIZE - 1;
3485         int ret;
3486         int nr = 0;
3487         size_t pg_offset = 0;
3488         loff_t i_size = i_size_read(inode);
3489         unsigned long end_index = i_size >> PAGE_SHIFT;
3490         int write_flags = 0;
3491         unsigned long nr_written = 0;
3492
3493         if (wbc->sync_mode == WB_SYNC_ALL)
3494                 write_flags = REQ_SYNC;
3495
3496         trace___extent_writepage(page, inode, wbc);
3497
3498         WARN_ON(!PageLocked(page));
3499
3500         ClearPageError(page);
3501
3502         pg_offset = i_size & (PAGE_SIZE - 1);
3503         if (page->index > end_index ||
3504            (page->index == end_index && !pg_offset)) {
3505                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3506                 unlock_page(page);
3507                 return 0;
3508         }
3509
3510         if (page->index == end_index) {
3511                 char *userpage;
3512
3513                 userpage = kmap_atomic(page);
3514                 memset(userpage + pg_offset, 0,
3515                        PAGE_SIZE - pg_offset);
3516                 kunmap_atomic(userpage);
3517                 flush_dcache_page(page);
3518         }
3519
3520         pg_offset = 0;
3521
3522         set_page_extent_mapped(page);
3523
3524         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3525         if (ret == 1)
3526                 goto done_unlocked;
3527         if (ret)
3528                 goto done;
3529
3530         ret = __extent_writepage_io(inode, page, wbc, epd,
3531                                     i_size, nr_written, write_flags, &nr);
3532         if (ret == 1)
3533                 goto done_unlocked;
3534
3535 done:
3536         if (nr == 0) {
3537                 /* make sure the mapping tag for page dirty gets cleared */
3538                 set_page_writeback(page);
3539                 end_page_writeback(page);
3540         }
3541         if (PageError(page)) {
3542                 ret = ret < 0 ? ret : -EIO;
3543                 end_extent_writepage(page, ret, start, page_end);
3544         }
3545         unlock_page(page);
3546         return ret;
3547
3548 done_unlocked:
3549         return 0;
3550 }
3551
3552 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3553 {
3554         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3555                        TASK_UNINTERRUPTIBLE);
3556 }
3557
3558 static noinline_for_stack int
3559 lock_extent_buffer_for_io(struct extent_buffer *eb,
3560                           struct btrfs_fs_info *fs_info,
3561                           struct extent_page_data *epd)
3562 {
3563         unsigned long i, num_pages;
3564         int flush = 0;
3565         int ret = 0;
3566
3567         if (!btrfs_try_tree_write_lock(eb)) {
3568                 flush = 1;
3569                 flush_write_bio(epd);
3570                 btrfs_tree_lock(eb);
3571         }
3572
3573         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3574                 btrfs_tree_unlock(eb);
3575                 if (!epd->sync_io)
3576                         return 0;
3577                 if (!flush) {
3578                         flush_write_bio(epd);
3579                         flush = 1;
3580                 }
3581                 while (1) {
3582                         wait_on_extent_buffer_writeback(eb);
3583                         btrfs_tree_lock(eb);
3584                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3585                                 break;
3586                         btrfs_tree_unlock(eb);
3587                 }
3588         }
3589
3590         /*
3591          * We need to do this to prevent races in people who check if the eb is
3592          * under IO since we can end up having no IO bits set for a short period
3593          * of time.
3594          */
3595         spin_lock(&eb->refs_lock);
3596         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3597                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3598                 spin_unlock(&eb->refs_lock);
3599                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3600                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3601                                      -eb->len,
3602                                      fs_info->dirty_metadata_batch);
3603                 ret = 1;
3604         } else {
3605                 spin_unlock(&eb->refs_lock);
3606         }
3607
3608         btrfs_tree_unlock(eb);
3609
3610         if (!ret)
3611                 return ret;
3612
3613         num_pages = num_extent_pages(eb->start, eb->len);
3614         for (i = 0; i < num_pages; i++) {
3615                 struct page *p = eb->pages[i];
3616
3617                 if (!trylock_page(p)) {
3618                         if (!flush) {
3619                                 flush_write_bio(epd);
3620                                 flush = 1;
3621                         }
3622                         lock_page(p);
3623                 }
3624         }
3625
3626         return ret;
3627 }
3628
3629 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3630 {
3631         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3632         smp_mb__after_atomic();
3633         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3634 }
3635
3636 static void set_btree_ioerr(struct page *page)
3637 {
3638         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3639
3640         SetPageError(page);
3641         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3642                 return;
3643
3644         /*
3645          * If writeback for a btree extent that doesn't belong to a log tree
3646          * failed, increment the counter transaction->eb_write_errors.
3647          * We do this because while the transaction is running and before it's
3648          * committing (when we call filemap_fdata[write|wait]_range against
3649          * the btree inode), we might have
3650          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3651          * returns an error or an error happens during writeback, when we're
3652          * committing the transaction we wouldn't know about it, since the pages
3653          * can be no longer dirty nor marked anymore for writeback (if a
3654          * subsequent modification to the extent buffer didn't happen before the
3655          * transaction commit), which makes filemap_fdata[write|wait]_range not
3656          * able to find the pages tagged with SetPageError at transaction
3657          * commit time. So if this happens we must abort the transaction,
3658          * otherwise we commit a super block with btree roots that point to
3659          * btree nodes/leafs whose content on disk is invalid - either garbage
3660          * or the content of some node/leaf from a past generation that got
3661          * cowed or deleted and is no longer valid.
3662          *
3663          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3664          * not be enough - we need to distinguish between log tree extents vs
3665          * non-log tree extents, and the next filemap_fdatawait_range() call
3666          * will catch and clear such errors in the mapping - and that call might
3667          * be from a log sync and not from a transaction commit. Also, checking
3668          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3669          * not done and would not be reliable - the eb might have been released
3670          * from memory and reading it back again means that flag would not be
3671          * set (since it's a runtime flag, not persisted on disk).
3672          *
3673          * Using the flags below in the btree inode also makes us achieve the
3674          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3675          * writeback for all dirty pages and before filemap_fdatawait_range()
3676          * is called, the writeback for all dirty pages had already finished
3677          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3678          * filemap_fdatawait_range() would return success, as it could not know
3679          * that writeback errors happened (the pages were no longer tagged for
3680          * writeback).
3681          */
3682         switch (eb->log_index) {
3683         case -1:
3684                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3685                 break;
3686         case 0:
3687                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3688                 break;
3689         case 1:
3690                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3691                 break;
3692         default:
3693                 BUG(); /* unexpected, logic error */
3694         }
3695 }
3696
3697 static void end_bio_extent_buffer_writepage(struct bio *bio)
3698 {
3699         struct bio_vec *bvec;
3700         struct extent_buffer *eb;
3701         int i, done;
3702
3703         bio_for_each_segment_all(bvec, bio, i) {
3704                 struct page *page = bvec->bv_page;
3705
3706                 eb = (struct extent_buffer *)page->private;
3707                 BUG_ON(!eb);
3708                 done = atomic_dec_and_test(&eb->io_pages);
3709
3710                 if (bio->bi_error ||
3711                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3712                         ClearPageUptodate(page);
3713                         set_btree_ioerr(page);
3714                 }
3715
3716                 end_page_writeback(page);
3717
3718                 if (!done)
3719                         continue;
3720
3721                 end_extent_buffer_writeback(eb);
3722         }
3723
3724         bio_put(bio);
3725 }
3726
3727 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3728                         struct btrfs_fs_info *fs_info,
3729                         struct writeback_control *wbc,
3730                         struct extent_page_data *epd)
3731 {
3732         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3733         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3734         u64 offset = eb->start;
3735         u32 nritems;
3736         unsigned long i, num_pages;
3737         unsigned long bio_flags = 0;
3738         unsigned long start, end;
3739         int write_flags = (epd->sync_io ? REQ_SYNC : 0) | REQ_META;
3740         int ret = 0;
3741
3742         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3743         num_pages = num_extent_pages(eb->start, eb->len);
3744         atomic_set(&eb->io_pages, num_pages);
3745         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3746                 bio_flags = EXTENT_BIO_TREE_LOG;
3747
3748         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3749         nritems = btrfs_header_nritems(eb);
3750         if (btrfs_header_level(eb) > 0) {
3751                 end = btrfs_node_key_ptr_offset(nritems);
3752
3753                 memzero_extent_buffer(eb, end, eb->len - end);
3754         } else {
3755                 /*
3756                  * leaf:
3757                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3758                  */
3759                 start = btrfs_item_nr_offset(nritems);
3760                 end = btrfs_leaf_data(eb) + leaf_data_end(fs_info, eb);
3761                 memzero_extent_buffer(eb, start, end - start);
3762         }
3763
3764         for (i = 0; i < num_pages; i++) {
3765                 struct page *p = eb->pages[i];
3766
3767                 clear_page_dirty_for_io(p);
3768                 set_page_writeback(p);
3769                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3770                                          p, offset >> 9, PAGE_SIZE, 0, bdev,
3771                                          &epd->bio,
3772                                          end_bio_extent_buffer_writepage,
3773                                          0, epd->bio_flags, bio_flags, false);
3774                 epd->bio_flags = bio_flags;
3775                 if (ret) {
3776                         set_btree_ioerr(p);
3777                         if (PageWriteback(p))
3778                                 end_page_writeback(p);
3779                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3780                                 end_extent_buffer_writeback(eb);
3781                         ret = -EIO;
3782                         break;
3783                 }
3784                 offset += PAGE_SIZE;
3785                 update_nr_written(wbc, 1);
3786                 unlock_page(p);
3787         }
3788
3789         if (unlikely(ret)) {
3790                 for (; i < num_pages; i++) {
3791                         struct page *p = eb->pages[i];
3792                         clear_page_dirty_for_io(p);
3793                         unlock_page(p);
3794                 }
3795         }
3796
3797         return ret;
3798 }
3799
3800 int btree_write_cache_pages(struct address_space *mapping,
3801                                    struct writeback_control *wbc)
3802 {
3803         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3804         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3805         struct extent_buffer *eb, *prev_eb = NULL;
3806         struct extent_page_data epd = {
3807                 .bio = NULL,
3808                 .tree = tree,
3809                 .extent_locked = 0,
3810                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3811                 .bio_flags = 0,
3812         };
3813         int ret = 0;
3814         int done = 0;
3815         int nr_to_write_done = 0;
3816         struct pagevec pvec;
3817         int nr_pages;
3818         pgoff_t index;
3819         pgoff_t end;            /* Inclusive */
3820         int scanned = 0;
3821         int tag;
3822
3823         pagevec_init(&pvec, 0);
3824         if (wbc->range_cyclic) {
3825                 index = mapping->writeback_index; /* Start from prev offset */
3826                 end = -1;
3827         } else {
3828                 index = wbc->range_start >> PAGE_SHIFT;
3829                 end = wbc->range_end >> PAGE_SHIFT;
3830                 scanned = 1;
3831         }
3832         if (wbc->sync_mode == WB_SYNC_ALL)
3833                 tag = PAGECACHE_TAG_TOWRITE;
3834         else
3835                 tag = PAGECACHE_TAG_DIRTY;
3836 retry:
3837         if (wbc->sync_mode == WB_SYNC_ALL)
3838                 tag_pages_for_writeback(mapping, index, end);
3839         while (!done && !nr_to_write_done && (index <= end) &&
3840                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3841                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3842                 unsigned i;
3843
3844                 scanned = 1;
3845                 for (i = 0; i < nr_pages; i++) {
3846                         struct page *page = pvec.pages[i];
3847
3848                         if (!PagePrivate(page))
3849                                 continue;
3850
3851                         if (!wbc->range_cyclic && page->index > end) {
3852                                 done = 1;
3853                                 break;
3854                         }
3855
3856                         spin_lock(&mapping->private_lock);
3857                         if (!PagePrivate(page)) {
3858                                 spin_unlock(&mapping->private_lock);
3859                                 continue;
3860                         }
3861
3862                         eb = (struct extent_buffer *)page->private;
3863
3864                         /*
3865                          * Shouldn't happen and normally this would be a BUG_ON
3866                          * but no sense in crashing the users box for something
3867                          * we can survive anyway.
3868                          */
3869                         if (WARN_ON(!eb)) {
3870                                 spin_unlock(&mapping->private_lock);
3871                                 continue;
3872                         }
3873
3874                         if (eb == prev_eb) {
3875                                 spin_unlock(&mapping->private_lock);
3876                                 continue;
3877                         }
3878
3879                         ret = atomic_inc_not_zero(&eb->refs);
3880                         spin_unlock(&mapping->private_lock);
3881                         if (!ret)
3882                                 continue;
3883
3884                         prev_eb = eb;
3885                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3886                         if (!ret) {
3887                                 free_extent_buffer(eb);
3888                                 continue;
3889                         }
3890
3891                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3892                         if (ret) {
3893                                 done = 1;
3894                                 free_extent_buffer(eb);
3895                                 break;
3896                         }
3897                         free_extent_buffer(eb);
3898
3899                         /*
3900                          * the filesystem may choose to bump up nr_to_write.
3901                          * We have to make sure to honor the new nr_to_write
3902                          * at any time
3903                          */
3904                         nr_to_write_done = wbc->nr_to_write <= 0;
3905                 }
3906                 pagevec_release(&pvec);
3907                 cond_resched();
3908         }
3909         if (!scanned && !done) {
3910                 /*
3911                  * We hit the last page and there is more work to be done: wrap
3912                  * back to the start of the file
3913                  */
3914                 scanned = 1;
3915                 index = 0;
3916                 goto retry;
3917         }
3918         flush_write_bio(&epd);
3919         return ret;
3920 }
3921
3922 /**
3923  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3924  * @mapping: address space structure to write
3925  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3926  * @writepage: function called for each page
3927  * @data: data passed to writepage function
3928  *
3929  * If a page is already under I/O, write_cache_pages() skips it, even
3930  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3931  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3932  * and msync() need to guarantee that all the data which was dirty at the time
3933  * the call was made get new I/O started against them.  If wbc->sync_mode is
3934  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3935  * existing IO to complete.
3936  */
3937 static int extent_write_cache_pages(struct address_space *mapping,
3938                              struct writeback_control *wbc,
3939                              writepage_t writepage, void *data,
3940                              void (*flush_fn)(void *))
3941 {
3942         struct inode *inode = mapping->host;
3943         int ret = 0;
3944         int done = 0;
3945         int nr_to_write_done = 0;
3946         struct pagevec pvec;
3947         int nr_pages;
3948         pgoff_t index;
3949         pgoff_t end;            /* Inclusive */
3950         pgoff_t done_index;
3951         int range_whole = 0;
3952         int scanned = 0;
3953         int tag;
3954
3955         /*
3956          * We have to hold onto the inode so that ordered extents can do their
3957          * work when the IO finishes.  The alternative to this is failing to add
3958          * an ordered extent if the igrab() fails there and that is a huge pain
3959          * to deal with, so instead just hold onto the inode throughout the
3960          * writepages operation.  If it fails here we are freeing up the inode
3961          * anyway and we'd rather not waste our time writing out stuff that is
3962          * going to be truncated anyway.
3963          */
3964         if (!igrab(inode))
3965                 return 0;
3966
3967         pagevec_init(&pvec, 0);
3968         if (wbc->range_cyclic) {
3969                 index = mapping->writeback_index; /* Start from prev offset */
3970                 end = -1;
3971         } else {
3972                 index = wbc->range_start >> PAGE_SHIFT;
3973                 end = wbc->range_end >> PAGE_SHIFT;
3974                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3975                         range_whole = 1;
3976                 scanned = 1;
3977         }
3978         if (wbc->sync_mode == WB_SYNC_ALL)
3979                 tag = PAGECACHE_TAG_TOWRITE;
3980         else
3981                 tag = PAGECACHE_TAG_DIRTY;
3982 retry:
3983         if (wbc->sync_mode == WB_SYNC_ALL)
3984                 tag_pages_for_writeback(mapping, index, end);
3985         done_index = index;
3986         while (!done && !nr_to_write_done && (index <= end) &&
3987                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3988                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3989                 unsigned i;
3990
3991                 scanned = 1;
3992                 for (i = 0; i < nr_pages; i++) {
3993                         struct page *page = pvec.pages[i];
3994
3995                         done_index = page->index;
3996                         /*
3997                          * At this point we hold neither mapping->tree_lock nor
3998                          * lock on the page itself: the page may be truncated or
3999                          * invalidated (changing page->mapping to NULL), or even
4000                          * swizzled back from swapper_space to tmpfs file
4001                          * mapping
4002                          */
4003                         if (!trylock_page(page)) {
4004                                 flush_fn(data);
4005                                 lock_page(page);
4006                         }
4007
4008                         if (unlikely(page->mapping != mapping)) {
4009                                 unlock_page(page);
4010                                 continue;
4011                         }
4012
4013                         if (!wbc->range_cyclic && page->index > end) {
4014                                 done = 1;
4015                                 unlock_page(page);
4016                                 continue;
4017                         }
4018
4019                         if (wbc->sync_mode != WB_SYNC_NONE) {
4020                                 if (PageWriteback(page))
4021                                         flush_fn(data);
4022                                 wait_on_page_writeback(page);
4023                         }
4024
4025                         if (PageWriteback(page) ||
4026                             !clear_page_dirty_for_io(page)) {
4027                                 unlock_page(page);
4028                                 continue;
4029                         }
4030
4031                         ret = (*writepage)(page, wbc, data);
4032
4033                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4034                                 unlock_page(page);
4035                                 ret = 0;
4036                         }
4037                         if (ret < 0) {
4038                                 /*
4039                                  * done_index is set past this page,
4040                                  * so media errors will not choke
4041                                  * background writeout for the entire
4042                                  * file. This has consequences for
4043                                  * range_cyclic semantics (ie. it may
4044                                  * not be suitable for data integrity
4045                                  * writeout).
4046                                  */
4047                                 done_index = page->index + 1;
4048                                 done = 1;
4049                                 break;
4050                         }
4051
4052                         /*
4053                          * the filesystem may choose to bump up nr_to_write.
4054                          * We have to make sure to honor the new nr_to_write
4055                          * at any time
4056                          */
4057                         nr_to_write_done = wbc->nr_to_write <= 0;
4058                 }
4059                 pagevec_release(&pvec);
4060                 cond_resched();
4061         }
4062         if (!scanned && !done) {
4063                 /*
4064                  * We hit the last page and there is more work to be done: wrap
4065                  * back to the start of the file
4066                  */
4067                 scanned = 1;
4068                 index = 0;
4069                 goto retry;
4070         }
4071
4072         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4073                 mapping->writeback_index = done_index;
4074
4075         btrfs_add_delayed_iput(inode);
4076         return ret;
4077 }
4078
4079 static void flush_epd_write_bio(struct extent_page_data *epd)
4080 {
4081         if (epd->bio) {
4082                 int ret;
4083
4084                 bio_set_op_attrs(epd->bio, REQ_OP_WRITE,
4085                                  epd->sync_io ? REQ_SYNC : 0);
4086
4087                 ret = submit_one_bio(epd->bio, 0, epd->bio_flags);
4088                 BUG_ON(ret < 0); /* -ENOMEM */
4089                 epd->bio = NULL;
4090         }
4091 }
4092
4093 static noinline void flush_write_bio(void *data)
4094 {
4095         struct extent_page_data *epd = data;
4096         flush_epd_write_bio(epd);
4097 }
4098
4099 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4100                           get_extent_t *get_extent,
4101                           struct writeback_control *wbc)
4102 {
4103         int ret;
4104         struct extent_page_data epd = {
4105                 .bio = NULL,
4106                 .tree = tree,
4107                 .get_extent = get_extent,
4108                 .extent_locked = 0,
4109                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4110                 .bio_flags = 0,
4111         };
4112
4113         ret = __extent_writepage(page, wbc, &epd);
4114
4115         flush_epd_write_bio(&epd);
4116         return ret;
4117 }
4118
4119 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4120                               u64 start, u64 end, get_extent_t *get_extent,
4121                               int mode)
4122 {
4123         int ret = 0;
4124         struct address_space *mapping = inode->i_mapping;
4125         struct page *page;
4126         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4127                 PAGE_SHIFT;
4128
4129         struct extent_page_data epd = {
4130                 .bio = NULL,
4131                 .tree = tree,
4132                 .get_extent = get_extent,
4133                 .extent_locked = 1,
4134                 .sync_io = mode == WB_SYNC_ALL,
4135                 .bio_flags = 0,
4136         };
4137         struct writeback_control wbc_writepages = {
4138                 .sync_mode      = mode,
4139                 .nr_to_write    = nr_pages * 2,
4140                 .range_start    = start,
4141                 .range_end      = end + 1,
4142         };
4143
4144         while (start <= end) {
4145                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4146                 if (clear_page_dirty_for_io(page))
4147                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4148                 else {
4149                         if (tree->ops && tree->ops->writepage_end_io_hook)
4150                                 tree->ops->writepage_end_io_hook(page, start,
4151                                                  start + PAGE_SIZE - 1,
4152                                                  NULL, 1);
4153                         unlock_page(page);
4154                 }
4155                 put_page(page);
4156                 start += PAGE_SIZE;
4157         }
4158
4159         flush_epd_write_bio(&epd);
4160         return ret;
4161 }
4162
4163 int extent_writepages(struct extent_io_tree *tree,
4164                       struct address_space *mapping,
4165                       get_extent_t *get_extent,
4166                       struct writeback_control *wbc)
4167 {
4168         int ret = 0;
4169         struct extent_page_data epd = {
4170                 .bio = NULL,
4171                 .tree = tree,
4172                 .get_extent = get_extent,
4173                 .extent_locked = 0,
4174                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4175                 .bio_flags = 0,
4176         };
4177
4178         ret = extent_write_cache_pages(mapping, wbc, __extent_writepage, &epd,
4179                                        flush_write_bio);
4180         flush_epd_write_bio(&epd);
4181         return ret;
4182 }
4183
4184 int extent_readpages(struct extent_io_tree *tree,
4185                      struct address_space *mapping,
4186                      struct list_head *pages, unsigned nr_pages,
4187                      get_extent_t get_extent)
4188 {
4189         struct bio *bio = NULL;
4190         unsigned page_idx;
4191         unsigned long bio_flags = 0;
4192         struct page *pagepool[16];
4193         struct page *page;
4194         struct extent_map *em_cached = NULL;
4195         int nr = 0;
4196         u64 prev_em_start = (u64)-1;
4197
4198         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4199                 page = list_entry(pages->prev, struct page, lru);
4200
4201                 prefetchw(&page->flags);
4202                 list_del(&page->lru);
4203                 if (add_to_page_cache_lru(page, mapping,
4204                                         page->index,
4205                                         readahead_gfp_mask(mapping))) {
4206                         put_page(page);
4207                         continue;
4208                 }
4209
4210                 pagepool[nr++] = page;
4211                 if (nr < ARRAY_SIZE(pagepool))
4212                         continue;
4213                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4214                                    &bio, 0, &bio_flags, &prev_em_start);
4215                 nr = 0;
4216         }
4217         if (nr)
4218                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4219                                    &bio, 0, &bio_flags, &prev_em_start);
4220
4221         if (em_cached)
4222                 free_extent_map(em_cached);
4223
4224         BUG_ON(!list_empty(pages));
4225         if (bio)
4226                 return submit_one_bio(bio, 0, bio_flags);
4227         return 0;
4228 }
4229
4230 /*
4231  * basic invalidatepage code, this waits on any locked or writeback
4232  * ranges corresponding to the page, and then deletes any extent state
4233  * records from the tree
4234  */
4235 int extent_invalidatepage(struct extent_io_tree *tree,
4236                           struct page *page, unsigned long offset)
4237 {
4238         struct extent_state *cached_state = NULL;
4239         u64 start = page_offset(page);
4240         u64 end = start + PAGE_SIZE - 1;
4241         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4242
4243         start += ALIGN(offset, blocksize);
4244         if (start > end)
4245                 return 0;
4246
4247         lock_extent_bits(tree, start, end, &cached_state);
4248         wait_on_page_writeback(page);
4249         clear_extent_bit(tree, start, end,
4250                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4251                          EXTENT_DO_ACCOUNTING,
4252                          1, 1, &cached_state, GFP_NOFS);
4253         return 0;
4254 }
4255
4256 /*
4257  * a helper for releasepage, this tests for areas of the page that
4258  * are locked or under IO and drops the related state bits if it is safe
4259  * to drop the page.
4260  */
4261 static int try_release_extent_state(struct extent_map_tree *map,
4262                                     struct extent_io_tree *tree,
4263                                     struct page *page, gfp_t mask)
4264 {
4265         u64 start = page_offset(page);
4266         u64 end = start + PAGE_SIZE - 1;
4267         int ret = 1;
4268
4269         if (test_range_bit(tree, start, end,
4270                            EXTENT_IOBITS, 0, NULL))
4271                 ret = 0;
4272         else {
4273                 /*
4274                  * at this point we can safely clear everything except the
4275                  * locked bit and the nodatasum bit
4276                  */
4277                 ret = clear_extent_bit(tree, start, end,
4278                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4279                                  0, 0, NULL, mask);
4280
4281                 /* if clear_extent_bit failed for enomem reasons,
4282                  * we can't allow the release to continue.
4283                  */
4284                 if (ret < 0)
4285                         ret = 0;
4286                 else
4287                         ret = 1;
4288         }
4289         return ret;
4290 }
4291
4292 /*
4293  * a helper for releasepage.  As long as there are no locked extents
4294  * in the range corresponding to the page, both state records and extent
4295  * map records are removed
4296  */
4297 int try_release_extent_mapping(struct extent_map_tree *map,
4298                                struct extent_io_tree *tree, struct page *page,
4299                                gfp_t mask)
4300 {
4301         struct extent_map *em;
4302         u64 start = page_offset(page);
4303         u64 end = start + PAGE_SIZE - 1;
4304
4305         if (gfpflags_allow_blocking(mask) &&
4306             page->mapping->host->i_size > SZ_16M) {
4307                 u64 len;
4308                 while (start <= end) {
4309                         len = end - start + 1;
4310                         write_lock(&map->lock);
4311                         em = lookup_extent_mapping(map, start, len);
4312                         if (!em) {
4313                                 write_unlock(&map->lock);
4314                                 break;
4315                         }
4316                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4317                             em->start != start) {
4318                                 write_unlock(&map->lock);
4319                                 free_extent_map(em);
4320                                 break;
4321                         }
4322                         if (!test_range_bit(tree, em->start,
4323                                             extent_map_end(em) - 1,
4324                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4325                                             0, NULL)) {
4326                                 remove_extent_mapping(map, em);
4327                                 /* once for the rb tree */
4328                                 free_extent_map(em);
4329                         }
4330                         start = extent_map_end(em);
4331                         write_unlock(&map->lock);
4332
4333                         /* once for us */
4334                         free_extent_map(em);
4335                 }
4336         }
4337         return try_release_extent_state(map, tree, page, mask);
4338 }
4339
4340 /*
4341  * helper function for fiemap, which doesn't want to see any holes.
4342  * This maps until we find something past 'last'
4343  */
4344 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4345                                                 u64 offset,
4346                                                 u64 last,
4347                                                 get_extent_t *get_extent)
4348 {
4349         u64 sectorsize = btrfs_inode_sectorsize(inode);
4350         struct extent_map *em;
4351         u64 len;
4352
4353         if (offset >= last)
4354                 return NULL;
4355
4356         while (1) {
4357                 len = last - offset;
4358                 if (len == 0)
4359                         break;
4360                 len = ALIGN(len, sectorsize);
4361                 em = get_extent(BTRFS_I(inode), NULL, 0, offset, len, 0);
4362                 if (IS_ERR_OR_NULL(em))
4363                         return em;
4364
4365                 /* if this isn't a hole return it */
4366                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4367                     em->block_start != EXTENT_MAP_HOLE) {
4368                         return em;
4369                 }
4370
4371                 /* this is a hole, advance to the next extent */
4372                 offset = extent_map_end(em);
4373                 free_extent_map(em);
4374                 if (offset >= last)
4375                         break;
4376         }
4377         return NULL;
4378 }
4379
4380 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4381                 __u64 start, __u64 len, get_extent_t *get_extent)
4382 {
4383         int ret = 0;
4384         u64 off = start;
4385         u64 max = start + len;
4386         u32 flags = 0;
4387         u32 found_type;
4388         u64 last;
4389         u64 last_for_get_extent = 0;
4390         u64 disko = 0;
4391         u64 isize = i_size_read(inode);
4392         struct btrfs_key found_key;
4393         struct extent_map *em = NULL;
4394         struct extent_state *cached_state = NULL;
4395         struct btrfs_path *path;
4396         struct btrfs_root *root = BTRFS_I(inode)->root;
4397         int end = 0;
4398         u64 em_start = 0;
4399         u64 em_len = 0;
4400         u64 em_end = 0;
4401
4402         if (len == 0)
4403                 return -EINVAL;
4404
4405         path = btrfs_alloc_path();
4406         if (!path)
4407                 return -ENOMEM;
4408         path->leave_spinning = 1;
4409
4410         start = round_down(start, btrfs_inode_sectorsize(inode));
4411         len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
4412
4413         /*
4414          * lookup the last file extent.  We're not using i_size here
4415          * because there might be preallocation past i_size
4416          */
4417         ret = btrfs_lookup_file_extent(NULL, root, path,
4418                         btrfs_ino(BTRFS_I(inode)), -1, 0);
4419         if (ret < 0) {
4420                 btrfs_free_path(path);
4421                 return ret;
4422         } else {
4423                 WARN_ON(!ret);
4424                 if (ret == 1)
4425                         ret = 0;
4426         }
4427
4428         path->slots[0]--;
4429         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4430         found_type = found_key.type;
4431
4432         /* No extents, but there might be delalloc bits */
4433         if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) ||
4434             found_type != BTRFS_EXTENT_DATA_KEY) {
4435                 /* have to trust i_size as the end */
4436                 last = (u64)-1;
4437                 last_for_get_extent = isize;
4438         } else {
4439                 /*
4440                  * remember the start of the last extent.  There are a
4441                  * bunch of different factors that go into the length of the
4442                  * extent, so its much less complex to remember where it started
4443                  */
4444                 last = found_key.offset;
4445                 last_for_get_extent = last + 1;
4446         }
4447         btrfs_release_path(path);
4448
4449         /*
4450          * we might have some extents allocated but more delalloc past those
4451          * extents.  so, we trust isize unless the start of the last extent is
4452          * beyond isize
4453          */
4454         if (last < isize) {
4455                 last = (u64)-1;
4456                 last_for_get_extent = isize;
4457         }
4458
4459         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4460                          &cached_state);
4461
4462         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4463                                    get_extent);
4464         if (!em)
4465                 goto out;
4466         if (IS_ERR(em)) {
4467                 ret = PTR_ERR(em);
4468                 goto out;
4469         }
4470
4471         while (!end) {
4472                 u64 offset_in_extent = 0;
4473
4474                 /* break if the extent we found is outside the range */
4475                 if (em->start >= max || extent_map_end(em) < off)
4476                         break;
4477
4478                 /*
4479                  * get_extent may return an extent that starts before our
4480                  * requested range.  We have to make sure the ranges
4481                  * we return to fiemap always move forward and don't
4482                  * overlap, so adjust the offsets here
4483                  */
4484                 em_start = max(em->start, off);
4485
4486                 /*
4487                  * record the offset from the start of the extent
4488                  * for adjusting the disk offset below.  Only do this if the
4489                  * extent isn't compressed since our in ram offset may be past
4490                  * what we have actually allocated on disk.
4491                  */
4492                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4493                         offset_in_extent = em_start - em->start;
4494                 em_end = extent_map_end(em);
4495                 em_len = em_end - em_start;
4496                 disko = 0;
4497                 flags = 0;
4498
4499                 /*
4500                  * bump off for our next call to get_extent
4501                  */
4502                 off = extent_map_end(em);
4503                 if (off >= max)
4504                         end = 1;
4505
4506                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4507                         end = 1;
4508                         flags |= FIEMAP_EXTENT_LAST;
4509                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4510                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4511                                   FIEMAP_EXTENT_NOT_ALIGNED);
4512                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4513                         flags |= (FIEMAP_EXTENT_DELALLOC |
4514                                   FIEMAP_EXTENT_UNKNOWN);
4515                 } else if (fieinfo->fi_extents_max) {
4516                         struct btrfs_trans_handle *trans;
4517
4518                         u64 bytenr = em->block_start -
4519                                 (em->start - em->orig_start);
4520
4521                         disko = em->block_start + offset_in_extent;
4522
4523                         /*
4524                          * We need a trans handle to get delayed refs
4525                          */
4526                         trans = btrfs_join_transaction(root);
4527                         /*
4528                          * It's OK if we can't start a trans we can still check
4529                          * from commit_root
4530                          */
4531                         if (IS_ERR(trans))
4532                                 trans = NULL;
4533
4534                         /*
4535                          * As btrfs supports shared space, this information
4536                          * can be exported to userspace tools via
4537                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4538                          * then we're just getting a count and we can skip the
4539                          * lookup stuff.
4540                          */
4541                         ret = btrfs_check_shared(trans, root->fs_info,
4542                                         root->objectid,
4543                                         btrfs_ino(BTRFS_I(inode)), bytenr);
4544                         if (trans)
4545                                 btrfs_end_transaction(trans);
4546                         if (ret < 0)
4547                                 goto out_free;
4548                         if (ret)
4549                                 flags |= FIEMAP_EXTENT_SHARED;
4550                         ret = 0;
4551                 }
4552                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4553                         flags |= FIEMAP_EXTENT_ENCODED;
4554                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4555                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4556
4557                 free_extent_map(em);
4558                 em = NULL;
4559                 if ((em_start >= last) || em_len == (u64)-1 ||
4560                    (last == (u64)-1 && isize <= em_end)) {
4561                         flags |= FIEMAP_EXTENT_LAST;
4562                         end = 1;
4563                 }
4564
4565                 /* now scan forward to see if this is really the last extent. */
4566                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4567                                            get_extent);
4568                 if (IS_ERR(em)) {
4569                         ret = PTR_ERR(em);
4570                         goto out;
4571                 }
4572                 if (!em) {
4573                         flags |= FIEMAP_EXTENT_LAST;
4574                         end = 1;
4575                 }
4576                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4577                                               em_len, flags);
4578                 if (ret) {
4579                         if (ret == 1)
4580                                 ret = 0;
4581                         goto out_free;
4582                 }
4583         }
4584 out_free:
4585         free_extent_map(em);
4586 out:
4587         btrfs_free_path(path);
4588         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4589                              &cached_state, GFP_NOFS);
4590         return ret;
4591 }
4592
4593 static void __free_extent_buffer(struct extent_buffer *eb)
4594 {
4595         btrfs_leak_debug_del(&eb->leak_list);
4596         kmem_cache_free(extent_buffer_cache, eb);
4597 }
4598
4599 int extent_buffer_under_io(struct extent_buffer *eb)
4600 {
4601         return (atomic_read(&eb->io_pages) ||
4602                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4603                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4604 }
4605
4606 /*
4607  * Helper for releasing extent buffer page.
4608  */
4609 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4610 {
4611         unsigned long index;
4612         struct page *page;
4613         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4614
4615         BUG_ON(extent_buffer_under_io(eb));
4616
4617         index = num_extent_pages(eb->start, eb->len);
4618         if (index == 0)
4619                 return;
4620
4621         do {
4622                 index--;
4623                 page = eb->pages[index];
4624                 if (!page)
4625                         continue;
4626                 if (mapped)
4627                         spin_lock(&page->mapping->private_lock);
4628                 /*
4629                  * We do this since we'll remove the pages after we've
4630                  * removed the eb from the radix tree, so we could race
4631                  * and have this page now attached to the new eb.  So
4632                  * only clear page_private if it's still connected to
4633                  * this eb.
4634                  */
4635                 if (PagePrivate(page) &&
4636                     page->private == (unsigned long)eb) {
4637                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4638                         BUG_ON(PageDirty(page));
4639                         BUG_ON(PageWriteback(page));
4640                         /*
4641                          * We need to make sure we haven't be attached
4642                          * to a new eb.
4643                          */
4644                         ClearPagePrivate(page);
4645                         set_page_private(page, 0);
4646                         /* One for the page private */
4647                         put_page(page);
4648                 }
4649
4650                 if (mapped)
4651                         spin_unlock(&page->mapping->private_lock);
4652
4653                 /* One for when we allocated the page */
4654                 put_page(page);
4655         } while (index != 0);
4656 }
4657
4658 /*
4659  * Helper for releasing the extent buffer.
4660  */
4661 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4662 {
4663         btrfs_release_extent_buffer_page(eb);
4664         __free_extent_buffer(eb);
4665 }
4666
4667 static struct extent_buffer *
4668 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4669                       unsigned long len)
4670 {
4671         struct extent_buffer *eb = NULL;
4672
4673         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4674         eb->start = start;
4675         eb->len = len;
4676         eb->fs_info = fs_info;
4677         eb->bflags = 0;
4678         rwlock_init(&eb->lock);
4679         atomic_set(&eb->write_locks, 0);
4680         atomic_set(&eb->read_locks, 0);
4681         atomic_set(&eb->blocking_readers, 0);
4682         atomic_set(&eb->blocking_writers, 0);
4683         atomic_set(&eb->spinning_readers, 0);
4684         atomic_set(&eb->spinning_writers, 0);
4685         eb->lock_nested = 0;
4686         init_waitqueue_head(&eb->write_lock_wq);
4687         init_waitqueue_head(&eb->read_lock_wq);
4688
4689         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4690
4691         spin_lock_init(&eb->refs_lock);
4692         atomic_set(&eb->refs, 1);
4693         atomic_set(&eb->io_pages, 0);
4694
4695         /*
4696          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4697          */
4698         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4699                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4700         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4701
4702         return eb;
4703 }
4704
4705 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4706 {
4707         unsigned long i;
4708         struct page *p;
4709         struct extent_buffer *new;
4710         unsigned long num_pages = num_extent_pages(src->start, src->len);
4711
4712         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4713         if (new == NULL)
4714                 return NULL;
4715
4716         for (i = 0; i < num_pages; i++) {
4717                 p = alloc_page(GFP_NOFS);
4718                 if (!p) {
4719                         btrfs_release_extent_buffer(new);
4720                         return NULL;
4721                 }
4722                 attach_extent_buffer_page(new, p);
4723                 WARN_ON(PageDirty(p));
4724                 SetPageUptodate(p);
4725                 new->pages[i] = p;
4726                 copy_page(page_address(p), page_address(src->pages[i]));
4727         }
4728
4729         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4730         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4731
4732         return new;
4733 }
4734
4735 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4736                                                   u64 start, unsigned long len)
4737 {
4738         struct extent_buffer *eb;
4739         unsigned long num_pages;
4740         unsigned long i;
4741
4742         num_pages = num_extent_pages(start, len);
4743
4744         eb = __alloc_extent_buffer(fs_info, start, len);
4745         if (!eb)
4746                 return NULL;
4747
4748         for (i = 0; i < num_pages; i++) {
4749                 eb->pages[i] = alloc_page(GFP_NOFS);
4750                 if (!eb->pages[i])
4751                         goto err;
4752         }
4753         set_extent_buffer_uptodate(eb);
4754         btrfs_set_header_nritems(eb, 0);
4755         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4756
4757         return eb;
4758 err:
4759         for (; i > 0; i--)
4760                 __free_page(eb->pages[i - 1]);
4761         __free_extent_buffer(eb);
4762         return NULL;
4763 }
4764
4765 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4766                                                 u64 start)
4767 {
4768         return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
4769 }
4770
4771 static void check_buffer_tree_ref(struct extent_buffer *eb)
4772 {
4773         int refs;
4774         /* the ref bit is tricky.  We have to make sure it is set
4775          * if we have the buffer dirty.   Otherwise the
4776          * code to free a buffer can end up dropping a dirty
4777          * page
4778          *
4779          * Once the ref bit is set, it won't go away while the
4780          * buffer is dirty or in writeback, and it also won't
4781          * go away while we have the reference count on the
4782          * eb bumped.
4783          *
4784          * We can't just set the ref bit without bumping the
4785          * ref on the eb because free_extent_buffer might
4786          * see the ref bit and try to clear it.  If this happens
4787          * free_extent_buffer might end up dropping our original
4788          * ref by mistake and freeing the page before we are able
4789          * to add one more ref.
4790          *
4791          * So bump the ref count first, then set the bit.  If someone
4792          * beat us to it, drop the ref we added.
4793          */
4794         refs = atomic_read(&eb->refs);
4795         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4796                 return;
4797
4798         spin_lock(&eb->refs_lock);
4799         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4800                 atomic_inc(&eb->refs);
4801         spin_unlock(&eb->refs_lock);
4802 }
4803
4804 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4805                 struct page *accessed)
4806 {
4807         unsigned long num_pages, i;
4808
4809         check_buffer_tree_ref(eb);
4810
4811         num_pages = num_extent_pages(eb->start, eb->len);
4812         for (i = 0; i < num_pages; i++) {
4813                 struct page *p = eb->pages[i];
4814
4815                 if (p != accessed)
4816                         mark_page_accessed(p);
4817         }
4818 }
4819
4820 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4821                                          u64 start)
4822 {
4823         struct extent_buffer *eb;
4824
4825         rcu_read_lock();
4826         eb = radix_tree_lookup(&fs_info->buffer_radix,
4827                                start >> PAGE_SHIFT);
4828         if (eb && atomic_inc_not_zero(&eb->refs)) {
4829                 rcu_read_unlock();
4830                 /*
4831                  * Lock our eb's refs_lock to avoid races with
4832                  * free_extent_buffer. When we get our eb it might be flagged
4833                  * with EXTENT_BUFFER_STALE and another task running
4834                  * free_extent_buffer might have seen that flag set,
4835                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4836                  * writeback flags not set) and it's still in the tree (flag
4837                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4838                  * of decrementing the extent buffer's reference count twice.
4839                  * So here we could race and increment the eb's reference count,
4840                  * clear its stale flag, mark it as dirty and drop our reference
4841                  * before the other task finishes executing free_extent_buffer,
4842                  * which would later result in an attempt to free an extent
4843                  * buffer that is dirty.
4844                  */
4845                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4846                         spin_lock(&eb->refs_lock);
4847                         spin_unlock(&eb->refs_lock);
4848                 }
4849                 mark_extent_buffer_accessed(eb, NULL);
4850                 return eb;
4851         }
4852         rcu_read_unlock();
4853
4854         return NULL;
4855 }
4856
4857 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4858 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4859                                         u64 start)
4860 {
4861         struct extent_buffer *eb, *exists = NULL;
4862         int ret;
4863
4864         eb = find_extent_buffer(fs_info, start);
4865         if (eb)
4866                 return eb;
4867         eb = alloc_dummy_extent_buffer(fs_info, start);
4868         if (!eb)
4869                 return NULL;
4870         eb->fs_info = fs_info;
4871 again:
4872         ret = radix_tree_preload(GFP_NOFS);
4873         if (ret)
4874                 goto free_eb;
4875         spin_lock(&fs_info->buffer_lock);
4876         ret = radix_tree_insert(&fs_info->buffer_radix,
4877                                 start >> PAGE_SHIFT, eb);
4878         spin_unlock(&fs_info->buffer_lock);
4879         radix_tree_preload_end();
4880         if (ret == -EEXIST) {
4881                 exists = find_extent_buffer(fs_info, start);
4882                 if (exists)
4883                         goto free_eb;
4884                 else
4885                         goto again;
4886         }
4887         check_buffer_tree_ref(eb);
4888         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4889
4890         /*
4891          * We will free dummy extent buffer's if they come into
4892          * free_extent_buffer with a ref count of 2, but if we are using this we
4893          * want the buffers to stay in memory until we're done with them, so
4894          * bump the ref count again.
4895          */
4896         atomic_inc(&eb->refs);
4897         return eb;
4898 free_eb:
4899         btrfs_release_extent_buffer(eb);
4900         return exists;
4901 }
4902 #endif
4903
4904 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4905                                           u64 start)
4906 {
4907         unsigned long len = fs_info->nodesize;
4908         unsigned long num_pages = num_extent_pages(start, len);
4909         unsigned long i;
4910         unsigned long index = start >> PAGE_SHIFT;
4911         struct extent_buffer *eb;
4912         struct extent_buffer *exists = NULL;
4913         struct page *p;
4914         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4915         int uptodate = 1;
4916         int ret;
4917
4918         if (!IS_ALIGNED(start, fs_info->sectorsize)) {
4919                 btrfs_err(fs_info, "bad tree block start %llu", start);
4920                 return ERR_PTR(-EINVAL);
4921         }
4922
4923         eb = find_extent_buffer(fs_info, start);
4924         if (eb)
4925                 return eb;
4926
4927         eb = __alloc_extent_buffer(fs_info, start, len);
4928         if (!eb)
4929                 return ERR_PTR(-ENOMEM);
4930
4931         for (i = 0; i < num_pages; i++, index++) {
4932                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4933                 if (!p) {
4934                         exists = ERR_PTR(-ENOMEM);
4935                         goto free_eb;
4936                 }
4937
4938                 spin_lock(&mapping->private_lock);
4939                 if (PagePrivate(p)) {
4940                         /*
4941                          * We could have already allocated an eb for this page
4942                          * and attached one so lets see if we can get a ref on
4943                          * the existing eb, and if we can we know it's good and
4944                          * we can just return that one, else we know we can just
4945                          * overwrite page->private.
4946                          */
4947                         exists = (struct extent_buffer *)p->private;
4948                         if (atomic_inc_not_zero(&exists->refs)) {
4949                                 spin_unlock(&mapping->private_lock);
4950                                 unlock_page(p);
4951                                 put_page(p);
4952                                 mark_extent_buffer_accessed(exists, p);
4953                                 goto free_eb;
4954                         }
4955                         exists = NULL;
4956
4957                         /*
4958                          * Do this so attach doesn't complain and we need to
4959                          * drop the ref the old guy had.
4960                          */
4961                         ClearPagePrivate(p);
4962                         WARN_ON(PageDirty(p));
4963                         put_page(p);
4964                 }
4965                 attach_extent_buffer_page(eb, p);
4966                 spin_unlock(&mapping->private_lock);
4967                 WARN_ON(PageDirty(p));
4968                 eb->pages[i] = p;
4969                 if (!PageUptodate(p))
4970                         uptodate = 0;
4971
4972                 /*
4973                  * see below about how we avoid a nasty race with release page
4974                  * and why we unlock later
4975                  */
4976         }
4977         if (uptodate)
4978                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4979 again:
4980         ret = radix_tree_preload(GFP_NOFS);
4981         if (ret) {
4982                 exists = ERR_PTR(ret);
4983                 goto free_eb;
4984         }
4985
4986         spin_lock(&fs_info->buffer_lock);
4987         ret = radix_tree_insert(&fs_info->buffer_radix,
4988                                 start >> PAGE_SHIFT, eb);
4989         spin_unlock(&fs_info->buffer_lock);
4990         radix_tree_preload_end();
4991         if (ret == -EEXIST) {
4992                 exists = find_extent_buffer(fs_info, start);
4993                 if (exists)
4994                         goto free_eb;
4995                 else
4996                         goto again;
4997         }
4998         /* add one reference for the tree */
4999         check_buffer_tree_ref(eb);
5000         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5001
5002         /*
5003          * there is a race where release page may have
5004          * tried to find this extent buffer in the radix
5005          * but failed.  It will tell the VM it is safe to
5006          * reclaim the, and it will clear the page private bit.
5007          * We must make sure to set the page private bit properly
5008          * after the extent buffer is in the radix tree so
5009          * it doesn't get lost
5010          */
5011         SetPageChecked(eb->pages[0]);
5012         for (i = 1; i < num_pages; i++) {
5013                 p = eb->pages[i];
5014                 ClearPageChecked(p);
5015                 unlock_page(p);
5016         }
5017         unlock_page(eb->pages[0]);
5018         return eb;
5019
5020 free_eb:
5021         WARN_ON(!atomic_dec_and_test(&eb->refs));
5022         for (i = 0; i < num_pages; i++) {
5023                 if (eb->pages[i])
5024                         unlock_page(eb->pages[i]);
5025         }
5026
5027         btrfs_release_extent_buffer(eb);
5028         return exists;
5029 }
5030
5031 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5032 {
5033         struct extent_buffer *eb =
5034                         container_of(head, struct extent_buffer, rcu_head);
5035
5036         __free_extent_buffer(eb);
5037 }
5038
5039 /* Expects to have eb->eb_lock already held */
5040 static int release_extent_buffer(struct extent_buffer *eb)
5041 {
5042         WARN_ON(atomic_read(&eb->refs) == 0);
5043         if (atomic_dec_and_test(&eb->refs)) {
5044                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5045                         struct btrfs_fs_info *fs_info = eb->fs_info;
5046
5047                         spin_unlock(&eb->refs_lock);
5048
5049                         spin_lock(&fs_info->buffer_lock);
5050                         radix_tree_delete(&fs_info->buffer_radix,
5051                                           eb->start >> PAGE_SHIFT);
5052                         spin_unlock(&fs_info->buffer_lock);
5053                 } else {
5054                         spin_unlock(&eb->refs_lock);
5055                 }
5056
5057                 /* Should be safe to release our pages at this point */
5058                 btrfs_release_extent_buffer_page(eb);
5059 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5060                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5061                         __free_extent_buffer(eb);
5062                         return 1;
5063                 }
5064 #endif
5065                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5066                 return 1;
5067         }
5068         spin_unlock(&eb->refs_lock);
5069
5070         return 0;
5071 }
5072
5073 void free_extent_buffer(struct extent_buffer *eb)
5074 {
5075         int refs;
5076         int old;
5077         if (!eb)
5078                 return;
5079
5080         while (1) {
5081                 refs = atomic_read(&eb->refs);
5082                 if (refs <= 3)
5083                         break;
5084                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5085                 if (old == refs)
5086                         return;
5087         }
5088
5089         spin_lock(&eb->refs_lock);
5090         if (atomic_read(&eb->refs) == 2 &&
5091             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5092                 atomic_dec(&eb->refs);
5093
5094         if (atomic_read(&eb->refs) == 2 &&
5095             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5096             !extent_buffer_under_io(eb) &&
5097             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5098                 atomic_dec(&eb->refs);
5099
5100         /*
5101          * I know this is terrible, but it's temporary until we stop tracking
5102          * the uptodate bits and such for the extent buffers.
5103          */
5104         release_extent_buffer(eb);
5105 }
5106
5107 void free_extent_buffer_stale(struct extent_buffer *eb)
5108 {
5109         if (!eb)
5110                 return;
5111
5112         spin_lock(&eb->refs_lock);
5113         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5114
5115         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5116             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5117                 atomic_dec(&eb->refs);
5118         release_extent_buffer(eb);
5119 }
5120
5121 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5122 {
5123         unsigned long i;
5124         unsigned long num_pages;
5125         struct page *page;
5126
5127         num_pages = num_extent_pages(eb->start, eb->len);
5128
5129         for (i = 0; i < num_pages; i++) {
5130                 page = eb->pages[i];
5131                 if (!PageDirty(page))
5132                         continue;
5133
5134                 lock_page(page);
5135                 WARN_ON(!PagePrivate(page));
5136
5137                 clear_page_dirty_for_io(page);
5138                 spin_lock_irq(&page->mapping->tree_lock);
5139                 if (!PageDirty(page)) {
5140                         radix_tree_tag_clear(&page->mapping->page_tree,
5141                                                 page_index(page),
5142                                                 PAGECACHE_TAG_DIRTY);
5143                 }
5144                 spin_unlock_irq(&page->mapping->tree_lock);
5145                 ClearPageError(page);
5146                 unlock_page(page);
5147         }
5148         WARN_ON(atomic_read(&eb->refs) == 0);
5149 }
5150
5151 int set_extent_buffer_dirty(struct extent_buffer *eb)
5152 {
5153         unsigned long i;
5154         unsigned long num_pages;
5155         int was_dirty = 0;
5156
5157         check_buffer_tree_ref(eb);
5158
5159         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5160
5161         num_pages = num_extent_pages(eb->start, eb->len);
5162         WARN_ON(atomic_read(&eb->refs) == 0);
5163         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5164
5165         for (i = 0; i < num_pages; i++)
5166                 set_page_dirty(eb->pages[i]);
5167         return was_dirty;
5168 }
5169
5170 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5171 {
5172         unsigned long i;
5173         struct page *page;
5174         unsigned long num_pages;
5175
5176         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5177         num_pages = num_extent_pages(eb->start, eb->len);
5178         for (i = 0; i < num_pages; i++) {
5179                 page = eb->pages[i];
5180                 if (page)
5181                         ClearPageUptodate(page);
5182         }
5183 }
5184
5185 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5186 {
5187         unsigned long i;
5188         struct page *page;
5189         unsigned long num_pages;
5190
5191         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5192         num_pages = num_extent_pages(eb->start, eb->len);
5193         for (i = 0; i < num_pages; i++) {
5194                 page = eb->pages[i];
5195                 SetPageUptodate(page);
5196         }
5197 }
5198
5199 int extent_buffer_uptodate(struct extent_buffer *eb)
5200 {
5201         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5202 }
5203
5204 int read_extent_buffer_pages(struct extent_io_tree *tree,
5205                              struct extent_buffer *eb, int wait,
5206                              get_extent_t *get_extent, int mirror_num)
5207 {
5208         unsigned long i;
5209         struct page *page;
5210         int err;
5211         int ret = 0;
5212         int locked_pages = 0;
5213         int all_uptodate = 1;
5214         unsigned long num_pages;
5215         unsigned long num_reads = 0;
5216         struct bio *bio = NULL;
5217         unsigned long bio_flags = 0;
5218
5219         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5220                 return 0;
5221
5222         num_pages = num_extent_pages(eb->start, eb->len);
5223         for (i = 0; i < num_pages; i++) {
5224                 page = eb->pages[i];
5225                 if (wait == WAIT_NONE) {
5226                         if (!trylock_page(page))
5227                                 goto unlock_exit;
5228                 } else {
5229                         lock_page(page);
5230                 }
5231                 locked_pages++;
5232         }
5233         /*
5234          * We need to firstly lock all pages to make sure that
5235          * the uptodate bit of our pages won't be affected by
5236          * clear_extent_buffer_uptodate().
5237          */
5238         for (i = 0; i < num_pages; i++) {
5239                 page = eb->pages[i];
5240                 if (!PageUptodate(page)) {
5241                         num_reads++;
5242                         all_uptodate = 0;
5243                 }
5244         }
5245
5246         if (all_uptodate) {
5247                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5248                 goto unlock_exit;
5249         }
5250
5251         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5252         eb->read_mirror = 0;
5253         atomic_set(&eb->io_pages, num_reads);
5254         for (i = 0; i < num_pages; i++) {
5255                 page = eb->pages[i];
5256
5257                 if (!PageUptodate(page)) {
5258                         if (ret) {
5259                                 atomic_dec(&eb->io_pages);
5260                                 unlock_page(page);
5261                                 continue;
5262                         }
5263
5264                         ClearPageError(page);
5265                         err = __extent_read_full_page(tree, page,
5266                                                       get_extent, &bio,
5267                                                       mirror_num, &bio_flags,
5268                                                       REQ_META);
5269                         if (err) {
5270                                 ret = err;
5271                                 /*
5272                                  * We use &bio in above __extent_read_full_page,
5273                                  * so we ensure that if it returns error, the
5274                                  * current page fails to add itself to bio and
5275                                  * it's been unlocked.
5276                                  *
5277                                  * We must dec io_pages by ourselves.
5278                                  */
5279                                 atomic_dec(&eb->io_pages);
5280                         }
5281                 } else {
5282                         unlock_page(page);
5283                 }
5284         }
5285
5286         if (bio) {
5287                 err = submit_one_bio(bio, mirror_num, bio_flags);
5288                 if (err)
5289                         return err;
5290         }
5291
5292         if (ret || wait != WAIT_COMPLETE)
5293                 return ret;
5294
5295         for (i = 0; i < num_pages; i++) {
5296                 page = eb->pages[i];
5297                 wait_on_page_locked(page);
5298                 if (!PageUptodate(page))
5299                         ret = -EIO;
5300         }
5301
5302         return ret;
5303
5304 unlock_exit:
5305         while (locked_pages > 0) {
5306                 locked_pages--;
5307                 page = eb->pages[locked_pages];
5308                 unlock_page(page);
5309         }
5310         return ret;
5311 }
5312
5313 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5314                         unsigned long start,
5315                         unsigned long len)
5316 {
5317         size_t cur;
5318         size_t offset;
5319         struct page *page;
5320         char *kaddr;
5321         char *dst = (char *)dstv;
5322         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5323         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5324
5325         WARN_ON(start > eb->len);
5326         WARN_ON(start + len > eb->start + eb->len);
5327
5328         offset = (start_offset + start) & (PAGE_SIZE - 1);
5329
5330         while (len > 0) {
5331                 page = eb->pages[i];
5332
5333                 cur = min(len, (PAGE_SIZE - offset));
5334                 kaddr = page_address(page);
5335                 memcpy(dst, kaddr + offset, cur);
5336
5337                 dst += cur;
5338                 len -= cur;
5339                 offset = 0;
5340                 i++;
5341         }
5342 }
5343
5344 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5345                         unsigned long start,
5346                         unsigned long len)
5347 {
5348         size_t cur;
5349         size_t offset;
5350         struct page *page;
5351         char *kaddr;
5352         char __user *dst = (char __user *)dstv;
5353         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5354         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5355         int ret = 0;
5356
5357         WARN_ON(start > eb->len);
5358         WARN_ON(start + len > eb->start + eb->len);
5359
5360         offset = (start_offset + start) & (PAGE_SIZE - 1);
5361
5362         while (len > 0) {
5363                 page = eb->pages[i];
5364
5365                 cur = min(len, (PAGE_SIZE - offset));
5366                 kaddr = page_address(page);
5367                 if (copy_to_user(dst, kaddr + offset, cur)) {
5368                         ret = -EFAULT;
5369                         break;
5370                 }
5371
5372                 dst += cur;
5373                 len -= cur;
5374                 offset = 0;
5375                 i++;
5376         }
5377
5378         return ret;
5379 }
5380
5381 /*
5382  * return 0 if the item is found within a page.
5383  * return 1 if the item spans two pages.
5384  * return -EINVAL otherwise.
5385  */
5386 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5387                                unsigned long min_len, char **map,
5388                                unsigned long *map_start,
5389                                unsigned long *map_len)
5390 {
5391         size_t offset = start & (PAGE_SIZE - 1);
5392         char *kaddr;
5393         struct page *p;
5394         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5395         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5396         unsigned long end_i = (start_offset + start + min_len - 1) >>
5397                 PAGE_SHIFT;
5398
5399         if (i != end_i)
5400                 return 1;
5401
5402         if (i == 0) {
5403                 offset = start_offset;
5404                 *map_start = 0;
5405         } else {
5406                 offset = 0;
5407                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5408         }
5409
5410         if (start + min_len > eb->len) {
5411                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5412                        eb->start, eb->len, start, min_len);
5413                 return -EINVAL;
5414         }
5415
5416         p = eb->pages[i];
5417         kaddr = page_address(p);
5418         *map = kaddr + offset;
5419         *map_len = PAGE_SIZE - offset;
5420         return 0;
5421 }
5422
5423 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5424                           unsigned long start,
5425                           unsigned long len)
5426 {
5427         size_t cur;
5428         size_t offset;
5429         struct page *page;
5430         char *kaddr;
5431         char *ptr = (char *)ptrv;
5432         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5433         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5434         int ret = 0;
5435
5436         WARN_ON(start > eb->len);
5437         WARN_ON(start + len > eb->start + eb->len);
5438
5439         offset = (start_offset + start) & (PAGE_SIZE - 1);
5440
5441         while (len > 0) {
5442                 page = eb->pages[i];
5443
5444                 cur = min(len, (PAGE_SIZE - offset));
5445
5446                 kaddr = page_address(page);
5447                 ret = memcmp(ptr, kaddr + offset, cur);
5448                 if (ret)
5449                         break;
5450
5451                 ptr += cur;
5452                 len -= cur;
5453                 offset = 0;
5454                 i++;
5455         }
5456         return ret;
5457 }
5458
5459 void write_extent_buffer_chunk_tree_uuid(struct extent_buffer *eb,
5460                 const void *srcv)
5461 {
5462         char *kaddr;
5463
5464         WARN_ON(!PageUptodate(eb->pages[0]));
5465         kaddr = page_address(eb->pages[0]);
5466         memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv,
5467                         BTRFS_FSID_SIZE);
5468 }
5469
5470 void write_extent_buffer_fsid(struct extent_buffer *eb, const void *srcv)
5471 {
5472         char *kaddr;
5473
5474         WARN_ON(!PageUptodate(eb->pages[0]));
5475         kaddr = page_address(eb->pages[0]);
5476         memcpy(kaddr + offsetof(struct btrfs_header, fsid), srcv,
5477                         BTRFS_FSID_SIZE);
5478 }
5479
5480 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5481                          unsigned long start, unsigned long len)
5482 {
5483         size_t cur;
5484         size_t offset;
5485         struct page *page;
5486         char *kaddr;
5487         char *src = (char *)srcv;
5488         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5489         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5490
5491         WARN_ON(start > eb->len);
5492         WARN_ON(start + len > eb->start + eb->len);
5493
5494         offset = (start_offset + start) & (PAGE_SIZE - 1);
5495
5496         while (len > 0) {
5497                 page = eb->pages[i];
5498                 WARN_ON(!PageUptodate(page));
5499
5500                 cur = min(len, PAGE_SIZE - offset);
5501                 kaddr = page_address(page);
5502                 memcpy(kaddr + offset, src, cur);
5503
5504                 src += cur;
5505                 len -= cur;
5506                 offset = 0;
5507                 i++;
5508         }
5509 }
5510
5511 void memzero_extent_buffer(struct extent_buffer *eb, unsigned long start,
5512                 unsigned long len)
5513 {
5514         size_t cur;
5515         size_t offset;
5516         struct page *page;
5517         char *kaddr;
5518         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5519         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5520
5521         WARN_ON(start > eb->len);
5522         WARN_ON(start + len > eb->start + eb->len);
5523
5524         offset = (start_offset + start) & (PAGE_SIZE - 1);
5525
5526         while (len > 0) {
5527                 page = eb->pages[i];
5528                 WARN_ON(!PageUptodate(page));
5529
5530                 cur = min(len, PAGE_SIZE - offset);
5531                 kaddr = page_address(page);
5532                 memset(kaddr + offset, 0, cur);
5533
5534                 len -= cur;
5535                 offset = 0;
5536                 i++;
5537         }
5538 }
5539
5540 void copy_extent_buffer_full(struct extent_buffer *dst,
5541                              struct extent_buffer *src)
5542 {
5543         int i;
5544         unsigned num_pages;
5545
5546         ASSERT(dst->len == src->len);
5547
5548         num_pages = num_extent_pages(dst->start, dst->len);
5549         for (i = 0; i < num_pages; i++)
5550                 copy_page(page_address(dst->pages[i]),
5551                                 page_address(src->pages[i]));
5552 }
5553
5554 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5555                         unsigned long dst_offset, unsigned long src_offset,
5556                         unsigned long len)
5557 {
5558         u64 dst_len = dst->len;
5559         size_t cur;
5560         size_t offset;
5561         struct page *page;
5562         char *kaddr;
5563         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5564         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5565
5566         WARN_ON(src->len != dst_len);
5567
5568         offset = (start_offset + dst_offset) &
5569                 (PAGE_SIZE - 1);
5570
5571         while (len > 0) {
5572                 page = dst->pages[i];
5573                 WARN_ON(!PageUptodate(page));
5574
5575                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5576
5577                 kaddr = page_address(page);
5578                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5579
5580                 src_offset += cur;
5581                 len -= cur;
5582                 offset = 0;
5583                 i++;
5584         }
5585 }
5586
5587 void le_bitmap_set(u8 *map, unsigned int start, int len)
5588 {
5589         u8 *p = map + BIT_BYTE(start);
5590         const unsigned int size = start + len;
5591         int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5592         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);
5593
5594         while (len - bits_to_set >= 0) {
5595                 *p |= mask_to_set;
5596                 len -= bits_to_set;
5597                 bits_to_set = BITS_PER_BYTE;
5598                 mask_to_set = ~0;
5599                 p++;
5600         }
5601         if (len) {
5602                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5603                 *p |= mask_to_set;
5604         }
5605 }
5606
5607 void le_bitmap_clear(u8 *map, unsigned int start, int len)
5608 {
5609         u8 *p = map + BIT_BYTE(start);
5610         const unsigned int size = start + len;
5611         int bits_to_clear = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5612         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(start);
5613
5614         while (len - bits_to_clear >= 0) {
5615                 *p &= ~mask_to_clear;
5616                 len -= bits_to_clear;
5617                 bits_to_clear = BITS_PER_BYTE;
5618                 mask_to_clear = ~0;
5619                 p++;
5620         }
5621         if (len) {
5622                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5623                 *p &= ~mask_to_clear;
5624         }
5625 }
5626
5627 /*
5628  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5629  * given bit number
5630  * @eb: the extent buffer
5631  * @start: offset of the bitmap item in the extent buffer
5632  * @nr: bit number
5633  * @page_index: return index of the page in the extent buffer that contains the
5634  * given bit number
5635  * @page_offset: return offset into the page given by page_index
5636  *
5637  * This helper hides the ugliness of finding the byte in an extent buffer which
5638  * contains a given bit.
5639  */
5640 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5641                                     unsigned long start, unsigned long nr,
5642                                     unsigned long *page_index,
5643                                     size_t *page_offset)
5644 {
5645         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5646         size_t byte_offset = BIT_BYTE(nr);
5647         size_t offset;
5648
5649         /*
5650          * The byte we want is the offset of the extent buffer + the offset of
5651          * the bitmap item in the extent buffer + the offset of the byte in the
5652          * bitmap item.
5653          */
5654         offset = start_offset + start + byte_offset;
5655
5656         *page_index = offset >> PAGE_SHIFT;
5657         *page_offset = offset & (PAGE_SIZE - 1);
5658 }
5659
5660 /**
5661  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5662  * @eb: the extent buffer
5663  * @start: offset of the bitmap item in the extent buffer
5664  * @nr: bit number to test
5665  */
5666 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5667                            unsigned long nr)
5668 {
5669         u8 *kaddr;
5670         struct page *page;
5671         unsigned long i;
5672         size_t offset;
5673
5674         eb_bitmap_offset(eb, start, nr, &i, &offset);
5675         page = eb->pages[i];
5676         WARN_ON(!PageUptodate(page));
5677         kaddr = page_address(page);
5678         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5679 }
5680
5681 /**
5682  * extent_buffer_bitmap_set - set an area of a bitmap
5683  * @eb: the extent buffer
5684  * @start: offset of the bitmap item in the extent buffer
5685  * @pos: bit number of the first bit
5686  * @len: number of bits to set
5687  */
5688 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5689                               unsigned long pos, unsigned long len)
5690 {
5691         u8 *kaddr;
5692         struct page *page;
5693         unsigned long i;
5694         size_t offset;
5695         const unsigned int size = pos + len;
5696         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5697         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5698
5699         eb_bitmap_offset(eb, start, pos, &i, &offset);
5700         page = eb->pages[i];
5701         WARN_ON(!PageUptodate(page));
5702         kaddr = page_address(page);
5703
5704         while (len >= bits_to_set) {
5705                 kaddr[offset] |= mask_to_set;
5706                 len -= bits_to_set;
5707                 bits_to_set = BITS_PER_BYTE;
5708                 mask_to_set = ~0;
5709                 if (++offset >= PAGE_SIZE && len > 0) {
5710                         offset = 0;
5711                         page = eb->pages[++i];
5712                         WARN_ON(!PageUptodate(page));
5713                         kaddr = page_address(page);
5714                 }
5715         }
5716         if (len) {
5717                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5718                 kaddr[offset] |= mask_to_set;
5719         }
5720 }
5721
5722
5723 /**
5724  * extent_buffer_bitmap_clear - clear an area of a bitmap
5725  * @eb: the extent buffer
5726  * @start: offset of the bitmap item in the extent buffer
5727  * @pos: bit number of the first bit
5728  * @len: number of bits to clear
5729  */
5730 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5731                                 unsigned long pos, unsigned long len)
5732 {
5733         u8 *kaddr;
5734         struct page *page;
5735         unsigned long i;
5736         size_t offset;
5737         const unsigned int size = pos + len;
5738         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5739         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5740
5741         eb_bitmap_offset(eb, start, pos, &i, &offset);
5742         page = eb->pages[i];
5743         WARN_ON(!PageUptodate(page));
5744         kaddr = page_address(page);
5745
5746         while (len >= bits_to_clear) {
5747                 kaddr[offset] &= ~mask_to_clear;
5748                 len -= bits_to_clear;
5749                 bits_to_clear = BITS_PER_BYTE;
5750                 mask_to_clear = ~0;
5751                 if (++offset >= PAGE_SIZE && len > 0) {
5752                         offset = 0;
5753                         page = eb->pages[++i];
5754                         WARN_ON(!PageUptodate(page));
5755                         kaddr = page_address(page);
5756                 }
5757         }
5758         if (len) {
5759                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5760                 kaddr[offset] &= ~mask_to_clear;
5761         }
5762 }
5763
5764 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5765 {
5766         unsigned long distance = (src > dst) ? src - dst : dst - src;
5767         return distance < len;
5768 }
5769
5770 static void copy_pages(struct page *dst_page, struct page *src_page,
5771                        unsigned long dst_off, unsigned long src_off,
5772                        unsigned long len)
5773 {
5774         char *dst_kaddr = page_address(dst_page);
5775         char *src_kaddr;
5776         int must_memmove = 0;
5777
5778         if (dst_page != src_page) {
5779                 src_kaddr = page_address(src_page);
5780         } else {
5781                 src_kaddr = dst_kaddr;
5782                 if (areas_overlap(src_off, dst_off, len))
5783                         must_memmove = 1;
5784         }
5785
5786         if (must_memmove)
5787                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5788         else
5789                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5790 }
5791
5792 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5793                            unsigned long src_offset, unsigned long len)
5794 {
5795         struct btrfs_fs_info *fs_info = dst->fs_info;
5796         size_t cur;
5797         size_t dst_off_in_page;
5798         size_t src_off_in_page;
5799         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5800         unsigned long dst_i;
5801         unsigned long src_i;
5802
5803         if (src_offset + len > dst->len) {
5804                 btrfs_err(fs_info,
5805                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5806                          src_offset, len, dst->len);
5807                 BUG_ON(1);
5808         }
5809         if (dst_offset + len > dst->len) {
5810                 btrfs_err(fs_info,
5811                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5812                          dst_offset, len, dst->len);
5813                 BUG_ON(1);
5814         }
5815
5816         while (len > 0) {
5817                 dst_off_in_page = (start_offset + dst_offset) &
5818                         (PAGE_SIZE - 1);
5819                 src_off_in_page = (start_offset + src_offset) &
5820                         (PAGE_SIZE - 1);
5821
5822                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5823                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5824
5825                 cur = min(len, (unsigned long)(PAGE_SIZE -
5826                                                src_off_in_page));
5827                 cur = min_t(unsigned long, cur,
5828                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5829
5830                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5831                            dst_off_in_page, src_off_in_page, cur);
5832
5833                 src_offset += cur;
5834                 dst_offset += cur;
5835                 len -= cur;
5836         }
5837 }
5838
5839 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5840                            unsigned long src_offset, unsigned long len)
5841 {
5842         struct btrfs_fs_info *fs_info = dst->fs_info;
5843         size_t cur;
5844         size_t dst_off_in_page;
5845         size_t src_off_in_page;
5846         unsigned long dst_end = dst_offset + len - 1;
5847         unsigned long src_end = src_offset + len - 1;
5848         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5849         unsigned long dst_i;
5850         unsigned long src_i;
5851
5852         if (src_offset + len > dst->len) {
5853                 btrfs_err(fs_info,
5854                           "memmove bogus src_offset %lu move len %lu len %lu",
5855                           src_offset, len, dst->len);
5856                 BUG_ON(1);
5857         }
5858         if (dst_offset + len > dst->len) {
5859                 btrfs_err(fs_info,
5860                           "memmove bogus dst_offset %lu move len %lu len %lu",
5861                           dst_offset, len, dst->len);
5862                 BUG_ON(1);
5863         }
5864         if (dst_offset < src_offset) {
5865                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5866                 return;
5867         }
5868         while (len > 0) {
5869                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5870                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5871
5872                 dst_off_in_page = (start_offset + dst_end) &
5873                         (PAGE_SIZE - 1);
5874                 src_off_in_page = (start_offset + src_end) &
5875                         (PAGE_SIZE - 1);
5876
5877                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5878                 cur = min(cur, dst_off_in_page + 1);
5879                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5880                            dst_off_in_page - cur + 1,
5881                            src_off_in_page - cur + 1, cur);
5882
5883                 dst_end -= cur;
5884                 src_end -= cur;
5885                 len -= cur;
5886         }
5887 }
5888
5889 int try_release_extent_buffer(struct page *page)
5890 {
5891         struct extent_buffer *eb;
5892
5893         /*
5894          * We need to make sure nobody is attaching this page to an eb right
5895          * now.
5896          */
5897         spin_lock(&page->mapping->private_lock);
5898         if (!PagePrivate(page)) {
5899                 spin_unlock(&page->mapping->private_lock);
5900                 return 1;
5901         }
5902
5903         eb = (struct extent_buffer *)page->private;
5904         BUG_ON(!eb);
5905
5906         /*
5907          * This is a little awful but should be ok, we need to make sure that
5908          * the eb doesn't disappear out from under us while we're looking at
5909          * this page.
5910          */
5911         spin_lock(&eb->refs_lock);
5912         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5913                 spin_unlock(&eb->refs_lock);
5914                 spin_unlock(&page->mapping->private_lock);
5915                 return 0;
5916         }
5917         spin_unlock(&page->mapping->private_lock);
5918
5919         /*
5920          * If tree ref isn't set then we know the ref on this eb is a real ref,
5921          * so just return, this page will likely be freed soon anyway.
5922          */
5923         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5924                 spin_unlock(&eb->refs_lock);
5925                 return 0;
5926         }
5927
5928         return release_extent_buffer(eb);
5929 }