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         if (tree->ops && tree->ops->check_extent_io_range)
  91                 tree->ops->check_extent_io_range(tree->private_data, caller,
  92                                                  start, end);
  93 }
  94 #else
  95 #define btrfs_leak_debug_add(new, head) do {} while (0)
  96 #define btrfs_leak_debug_del(entry)     do {} while (0)
  97 #define btrfs_leak_debug_check()        do {} while (0)
  98 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
  99 #endif
 100
 101 #define BUFFER_LRU_MAX 64
 102
 103 struct tree_entry {
 104         u64 start;
 105         u64 end;
 106         struct rb_node rb_node;
 107 };
 108
 109 struct extent_page_data {
 110         struct bio *bio;
 111         struct extent_io_tree *tree;
 112         get_extent_t *get_extent;
 113         unsigned long bio_flags;
 114
 115         /* tells writepage not to lock the state bits for this range
 116          * it still does the unlocking
 117          */
 118         unsigned int extent_locked:1;
 119
 120         /* tells the submit_bio code to use REQ_SYNC */
 121         unsigned int sync_io:1;
 122 };
 123
 124 static void add_extent_changeset(struct extent_state *state, unsigned bits,
 125                                  struct extent_changeset *changeset,
 126                                  int set)
 127 {
 128         int ret;
 129
 130         if (!changeset)
 131                 return;
 132         if (set && (state->state & bits) == bits)
 133                 return;
 134         if (!set && (state->state & bits) == 0)
 135                 return;
 136         changeset->bytes_changed += state->end - state->start + 1;
 137         ret = ulist_add(&changeset->range_changed, state->start, state->end,
 138                         GFP_ATOMIC);
 139         /* ENOMEM */
 140         BUG_ON(ret < 0);
 141 }
 142
 143 static noinline void flush_write_bio(void *data);
 144 static inline struct btrfs_fs_info *
 145 tree_fs_info(struct extent_io_tree *tree)
 146 {
 147         if (tree->ops)
 148                 return tree->ops->tree_fs_info(tree->private_data);
 149         return NULL;
 150 }
 151
 152 int __init extent_io_init(void)
 153 {
 154         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 155                         sizeof(struct extent_state), 0,
 156                         SLAB_MEM_SPREAD, NULL);
 157         if (!extent_state_cache)
 158                 return -ENOMEM;
 159
 160         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 161                         sizeof(struct extent_buffer), 0,
 162                         SLAB_MEM_SPREAD, NULL);
 163         if (!extent_buffer_cache)
 164                 goto free_state_cache;
 165
 166         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 167                                      offsetof(struct btrfs_io_bio, bio),
 168                                      BIOSET_NEED_BVECS);
 169         if (!btrfs_bioset)
 170                 goto free_buffer_cache;
 171
 172         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 173                 goto free_bioset;
 174
 175         return 0;
 176
 177 free_bioset:
 178         bioset_free(btrfs_bioset);
 179         btrfs_bioset = NULL;
 180
 181 free_buffer_cache:
 182         kmem_cache_destroy(extent_buffer_cache);
 183         extent_buffer_cache = NULL;
 184
 185 free_state_cache:
 186         kmem_cache_destroy(extent_state_cache);
 187         extent_state_cache = NULL;
 188         return -ENOMEM;
 189 }
 190
 191 void extent_io_exit(void)
 192 {
 193         btrfs_leak_debug_check();
 194
 195         /*
 196          * Make sure all delayed rcu free are flushed before we
 197          * destroy caches.
 198          */
 199         rcu_barrier();
 200         kmem_cache_destroy(extent_state_cache);
 201         kmem_cache_destroy(extent_buffer_cache);
 202         if (btrfs_bioset)
 203                 bioset_free(btrfs_bioset);
 204 }
 205
 206 void extent_io_tree_init(struct extent_io_tree *tree,
 207                          void *private_data)
 208 {
 209         tree->state = RB_ROOT;
 210         tree->ops = NULL;
 211         tree->dirty_bytes = 0;
 212         spin_lock_init(&tree->lock);
 213         tree->private_data = private_data;
 214 }
 215
 216 static struct extent_state *alloc_extent_state(gfp_t mask)
 217 {
 218         struct extent_state *state;
 219
 220         /*
 221          * The given mask might be not appropriate for the slab allocator,
 222          * drop the unsupported bits
 223          */
 224         mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
 225         state = kmem_cache_alloc(extent_state_cache, mask);
 226         if (!state)
 227                 return state;
 228         state->state = 0;
 229         state->failrec = NULL;
 230         RB_CLEAR_NODE(&state->rb_node);
 231         btrfs_leak_debug_add(&state->leak_list, &states);
 232         refcount_set(&state->refs, 1);
 233         init_waitqueue_head(&state->wq);
 234         trace_alloc_extent_state(state, mask, _RET_IP_);
 235         return state;
 236 }
 237
 238 void free_extent_state(struct extent_state *state)
 239 {
 240         if (!state)
 241                 return;
 242         if (refcount_dec_and_test(&state->refs)) {
 243                 WARN_ON(extent_state_in_tree(state));
 244                 btrfs_leak_debug_del(&state->leak_list);
 245                 trace_free_extent_state(state, _RET_IP_);
 246                 kmem_cache_free(extent_state_cache, state);
 247         }
 248 }
 249
 250 static struct rb_node *tree_insert(struct rb_root *root,
 251                                    struct rb_node *search_start,
 252                                    u64 offset,
 253                                    struct rb_node *node,
 254                                    struct rb_node ***p_in,
 255                                    struct rb_node **parent_in)
 256 {
 257         struct rb_node **p;
 258         struct rb_node *parent = NULL;
 259         struct tree_entry *entry;
 260
 261         if (p_in && parent_in) {
 262                 p = *p_in;
 263                 parent = *parent_in;
 264                 goto do_insert;
 265         }
 266
 267         p = search_start ? &search_start : &root->rb_node;
 268         while (*p) {
 269                 parent = *p;
 270                 entry = rb_entry(parent, struct tree_entry, rb_node);
 271
 272                 if (offset < entry->start)
 273                         p = &(*p)->rb_left;
 274                 else if (offset > entry->end)
 275                         p = &(*p)->rb_right;
 276                 else
 277                         return parent;
 278         }
 279
 280 do_insert:
 281         rb_link_node(node, parent, p);
 282         rb_insert_color(node, root);
 283         return NULL;
 284 }
 285
 286 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 287                                       struct rb_node **prev_ret,
 288                                       struct rb_node **next_ret,
 289                                       struct rb_node ***p_ret,
 290                                       struct rb_node **parent_ret)
 291 {
 292         struct rb_root *root = &tree->state;
 293         struct rb_node **n = &root->rb_node;
 294         struct rb_node *prev = NULL;
 295         struct rb_node *orig_prev = NULL;
 296         struct tree_entry *entry;
 297         struct tree_entry *prev_entry = NULL;
 298
 299         while (*n) {
 300                 prev = *n;
 301                 entry = rb_entry(prev, struct tree_entry, rb_node);
 302                 prev_entry = entry;
 303
 304                 if (offset < entry->start)
 305                         n = &(*n)->rb_left;
 306                 else if (offset > entry->end)
 307                         n = &(*n)->rb_right;
 308                 else
 309                         return *n;
 310         }
 311
 312         if (p_ret)
 313                 *p_ret = n;
 314         if (parent_ret)
 315                 *parent_ret = prev;
 316
 317         if (prev_ret) {
 318                 orig_prev = prev;
 319                 while (prev && offset > prev_entry->end) {
 320                         prev = rb_next(prev);
 321                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 322                 }
 323                 *prev_ret = prev;
 324                 prev = orig_prev;
 325         }
 326
 327         if (next_ret) {
 328                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 329                 while (prev && offset < prev_entry->start) {
 330                         prev = rb_prev(prev);
 331                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 332                 }
 333                 *next_ret = prev;
 334         }
 335         return NULL;
 336 }
 337
 338 static inline struct rb_node *
 339 tree_search_for_insert(struct extent_io_tree *tree,
 340                        u64 offset,
 341                        struct rb_node ***p_ret,
 342                        struct rb_node **parent_ret)
 343 {
 344         struct rb_node *prev = NULL;
 345         struct rb_node *ret;
 346
 347         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 348         if (!ret)
 349                 return prev;
 350         return ret;
 351 }
 352
 353 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 354                                           u64 offset)
 355 {
 356         return tree_search_for_insert(tree, offset, NULL, NULL);
 357 }
 358
 359 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 360                      struct extent_state *other)
 361 {
 362         if (tree->ops && tree->ops->merge_extent_hook)
 363                 tree->ops->merge_extent_hook(tree->private_data, new, other);
 364 }
 365
 366 /*
 367  * utility function to look for merge candidates inside a given range.
 368  * Any extents with matching state are merged together into a single
 369  * extent in the tree.  Extents with EXTENT_IO in their state field
 370  * are not merged because the end_io handlers need to be able to do
 371  * operations on them without sleeping (or doing allocations/splits).
 372  *
 373  * This should be called with the tree lock held.
 374  */
 375 static void merge_state(struct extent_io_tree *tree,
 376                         struct extent_state *state)
 377 {
 378         struct extent_state *other;
 379         struct rb_node *other_node;
 380
 381         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 382                 return;
 383
 384         other_node = rb_prev(&state->rb_node);
 385         if (other_node) {
 386                 other = rb_entry(other_node, struct extent_state, rb_node);
 387                 if (other->end == state->start - 1 &&
 388                     other->state == state->state) {
 389                         merge_cb(tree, state, other);
 390                         state->start = other->start;
 391                         rb_erase(&other->rb_node, &tree->state);
 392                         RB_CLEAR_NODE(&other->rb_node);
 393                         free_extent_state(other);
 394                 }
 395         }
 396         other_node = rb_next(&state->rb_node);
 397         if (other_node) {
 398                 other = rb_entry(other_node, struct extent_state, rb_node);
 399                 if (other->start == state->end + 1 &&
 400                     other->state == state->state) {
 401                         merge_cb(tree, state, other);
 402                         state->end = other->end;
 403                         rb_erase(&other->rb_node, &tree->state);
 404                         RB_CLEAR_NODE(&other->rb_node);
 405                         free_extent_state(other);
 406                 }
 407         }
 408 }
 409
 410 static void set_state_cb(struct extent_io_tree *tree,
 411                          struct extent_state *state, unsigned *bits)
 412 {
 413         if (tree->ops && tree->ops->set_bit_hook)
 414                 tree->ops->set_bit_hook(tree->private_data, state, bits);
 415 }
 416
 417 static void clear_state_cb(struct extent_io_tree *tree,
 418                            struct extent_state *state, unsigned *bits)
 419 {
 420         if (tree->ops && tree->ops->clear_bit_hook)
 421                 tree->ops->clear_bit_hook(tree->private_data, state, bits);
 422 }
 423
 424 static void set_state_bits(struct extent_io_tree *tree,
 425                            struct extent_state *state, unsigned *bits,
 426                            struct extent_changeset *changeset);
 427
 428 /*
 429  * insert an extent_state struct into the tree.  'bits' are set on the
 430  * struct before it is inserted.
 431  *
 432  * This may return -EEXIST if the extent is already there, in which case the
 433  * state struct is freed.
 434  *
 435  * The tree lock is not taken internally.  This is a utility function and
 436  * probably isn't what you want to call (see set/clear_extent_bit).
 437  */
 438 static int insert_state(struct extent_io_tree *tree,
 439                         struct extent_state *state, u64 start, u64 end,
 440                         struct rb_node ***p,
 441                         struct rb_node **parent,
 442                         unsigned *bits, struct extent_changeset *changeset)
 443 {
 444         struct rb_node *node;
 445
 446         if (end < start)
 447                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 448                        end, start);
 449         state->start = start;
 450         state->end = end;
 451
 452         set_state_bits(tree, state, bits, changeset);
 453
 454         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 455         if (node) {
 456                 struct extent_state *found;
 457                 found = rb_entry(node, struct extent_state, rb_node);
 458                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
 459                        found->start, found->end, start, end);
 460                 return -EEXIST;
 461         }
 462         merge_state(tree, state);
 463         return 0;
 464 }
 465
 466 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 467                      u64 split)
 468 {
 469         if (tree->ops && tree->ops->split_extent_hook)
 470                 tree->ops->split_extent_hook(tree->private_data, orig, split);
 471 }
 472
 473 /*
 474  * split a given extent state struct in two, inserting the preallocated
 475  * struct 'prealloc' as the newly created second half.  'split' indicates an
 476  * offset inside 'orig' where it should be split.
 477  *
 478  * Before calling,
 479  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 480  * are two extent state structs in the tree:
 481  * prealloc: [orig->start, split - 1]
 482  * orig: [ split, orig->end ]
 483  *
 484  * The tree locks are not taken by this function. They need to be held
 485  * by the caller.
 486  */
 487 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 488                        struct extent_state *prealloc, u64 split)
 489 {
 490         struct rb_node *node;
 491
 492         split_cb(tree, orig, split);
 493
 494         prealloc->start = orig->start;
 495         prealloc->end = split - 1;
 496         prealloc->state = orig->state;
 497         orig->start = split;
 498
 499         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 500                            &prealloc->rb_node, NULL, NULL);
 501         if (node) {
 502                 free_extent_state(prealloc);
 503                 return -EEXIST;
 504         }
 505         return 0;
 506 }
 507
 508 static struct extent_state *next_state(struct extent_state *state)
 509 {
 510         struct rb_node *next = rb_next(&state->rb_node);
 511         if (next)
 512                 return rb_entry(next, struct extent_state, rb_node);
 513         else
 514                 return NULL;
 515 }
 516
 517 /*
 518  * utility function to clear some bits in an extent state struct.
 519  * it will optionally wake up any one waiting on this state (wake == 1).
 520  *
 521  * If no bits are set on the state struct after clearing things, the
 522  * struct is freed and removed from the tree
 523  */
 524 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 525                                             struct extent_state *state,
 526                                             unsigned *bits, int wake,
 527                                             struct extent_changeset *changeset)
 528 {
 529         struct extent_state *next;
 530         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 531
 532         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 533                 u64 range = state->end - state->start + 1;
 534                 WARN_ON(range > tree->dirty_bytes);
 535                 tree->dirty_bytes -= range;
 536         }
 537         clear_state_cb(tree, state, bits);
 538         add_extent_changeset(state, bits_to_clear, changeset, 0);
 539         state->state &= ~bits_to_clear;
 540         if (wake)
 541                 wake_up(&state->wq);
 542         if (state->state == 0) {
 543                 next = next_state(state);
 544                 if (extent_state_in_tree(state)) {
 545                         rb_erase(&state->rb_node, &tree->state);
 546                         RB_CLEAR_NODE(&state->rb_node);
 547                         free_extent_state(state);
 548                 } else {
 549                         WARN_ON(1);
 550                 }
 551         } else {
 552                 merge_state(tree, state);
 553                 next = next_state(state);
 554         }
 555         return next;
 556 }
 557
 558 static struct extent_state *
 559 alloc_extent_state_atomic(struct extent_state *prealloc)
 560 {
 561         if (!prealloc)
 562                 prealloc = alloc_extent_state(GFP_ATOMIC);
 563
 564         return prealloc;
 565 }
 566
 567 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 568 {
 569         btrfs_panic(tree_fs_info(tree), err,
 570                     "Locking error: Extent tree was modified by another thread while locked.");
 571 }
 572
 573 /*
 574  * clear some bits on a range in the tree.  This may require splitting
 575  * or inserting elements in the tree, so the gfp mask is used to
 576  * indicate which allocations or sleeping are allowed.
 577  *
 578  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 579  * the given range from the tree regardless of state (ie for truncate).
 580  *
 581  * the range [start, end] is inclusive.
 582  *
 583  * This takes the tree lock, and returns 0 on success and < 0 on error.
 584  */
 585 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 586                               unsigned bits, int wake, int delete,
 587                               struct extent_state **cached_state,
 588                               gfp_t mask, struct extent_changeset *changeset)
 589 {
 590         struct extent_state *state;
 591         struct extent_state *cached;
 592         struct extent_state *prealloc = NULL;
 593         struct rb_node *node;
 594         u64 last_end;
 595         int err;
 596         int clear = 0;
 597
 598         btrfs_debug_check_extent_io_range(tree, start, end);
 599
 600         if (bits & EXTENT_DELALLOC)
 601                 bits |= EXTENT_NORESERVE;
 602
 603         if (delete)
 604                 bits |= ~EXTENT_CTLBITS;
 605         bits |= EXTENT_FIRST_DELALLOC;
 606
 607         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 608                 clear = 1;
 609 again:
 610         if (!prealloc && gfpflags_allow_blocking(mask)) {
 611                 /*
 612                  * Don't care for allocation failure here because we might end
 613                  * up not needing the pre-allocated extent state at all, which
 614                  * is the case if we only have in the tree extent states that
 615                  * cover our input range and don't cover too any other range.
 616                  * If we end up needing a new extent state we allocate it later.
 617                  */
 618                 prealloc = alloc_extent_state(mask);
 619         }
 620
 621         spin_lock(&tree->lock);
 622         if (cached_state) {
 623                 cached = *cached_state;
 624
 625                 if (clear) {
 626                         *cached_state = NULL;
 627                         cached_state = NULL;
 628                 }
 629
 630                 if (cached && extent_state_in_tree(cached) &&
 631                     cached->start <= start && cached->end > start) {
 632                         if (clear)
 633                                 refcount_dec(&cached->refs);
 634                         state = cached;
 635                         goto hit_next;
 636                 }
 637                 if (clear)
 638                         free_extent_state(cached);
 639         }
 640         /*
 641          * this search will find the extents that end after
 642          * our range starts
 643          */
 644         node = tree_search(tree, start);
 645         if (!node)
 646                 goto out;
 647         state = rb_entry(node, struct extent_state, rb_node);
 648 hit_next:
 649         if (state->start > end)
 650                 goto out;
 651         WARN_ON(state->end < start);
 652         last_end = state->end;
 653
 654         /* the state doesn't have the wanted bits, go ahead */
 655         if (!(state->state & bits)) {
 656                 state = next_state(state);
 657                 goto next;
 658         }
 659
 660         /*
 661          *     | ---- desired range ---- |
 662          *  | state | or
 663          *  | ------------- state -------------- |
 664          *
 665          * We need to split the extent we found, and may flip
 666          * bits on second half.
 667          *
 668          * If the extent we found extends past our range, we
 669          * just split and search again.  It'll get split again
 670          * the next time though.
 671          *
 672          * If the extent we found is inside our range, we clear
 673          * the desired bit on it.
 674          */
 675
 676         if (state->start < start) {
 677                 prealloc = alloc_extent_state_atomic(prealloc);
 678                 BUG_ON(!prealloc);
 679                 err = split_state(tree, state, prealloc, start);
 680                 if (err)
 681                         extent_io_tree_panic(tree, err);
 682
 683                 prealloc = NULL;
 684                 if (err)
 685                         goto out;
 686                 if (state->end <= end) {
 687                         state = clear_state_bit(tree, state, &bits, wake,
 688                                                 changeset);
 689                         goto next;
 690                 }
 691                 goto search_again;
 692         }
 693         /*
 694          * | ---- desired range ---- |
 695          *                        | state |
 696          * We need to split the extent, and clear the bit
 697          * on the first half
 698          */
 699         if (state->start <= end && state->end > end) {
 700                 prealloc = alloc_extent_state_atomic(prealloc);
 701                 BUG_ON(!prealloc);
 702                 err = split_state(tree, state, prealloc, end + 1);
 703                 if (err)
 704                         extent_io_tree_panic(tree, err);
 705
 706                 if (wake)
 707                         wake_up(&state->wq);
 708
 709                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 710
 711                 prealloc = NULL;
 712                 goto out;
 713         }
 714
 715         state = clear_state_bit(tree, state, &bits, wake, changeset);
 716 next:
 717         if (last_end == (u64)-1)
 718                 goto out;
 719         start = last_end + 1;
 720         if (start <= end && state && !need_resched())
 721                 goto hit_next;
 722
 723 search_again:
 724         if (start > end)
 725                 goto out;
 726         spin_unlock(&tree->lock);
 727         if (gfpflags_allow_blocking(mask))
 728                 cond_resched();
 729         goto again;
 730
 731 out:
 732         spin_unlock(&tree->lock);
 733         if (prealloc)
 734                 free_extent_state(prealloc);
 735
 736         return 0;
 737
 738 }
 739
 740 static void wait_on_state(struct extent_io_tree *tree,
 741                           struct extent_state *state)
 742                 __releases(tree->lock)
 743                 __acquires(tree->lock)
 744 {
 745         DEFINE_WAIT(wait);
 746         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 747         spin_unlock(&tree->lock);
 748         schedule();
 749         spin_lock(&tree->lock);
 750         finish_wait(&state->wq, &wait);
 751 }
 752
 753 /*
 754  * waits for one or more bits to clear on a range in the state tree.
 755  * The range [start, end] is inclusive.
 756  * The tree lock is taken by this function
 757  */
 758 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 759                             unsigned long bits)
 760 {
 761         struct extent_state *state;
 762         struct rb_node *node;
 763
 764         btrfs_debug_check_extent_io_range(tree, start, end);
 765
 766         spin_lock(&tree->lock);
 767 again:
 768         while (1) {
 769                 /*
 770                  * this search will find all the extents that end after
 771                  * our range starts
 772                  */
 773                 node = tree_search(tree, start);
 774 process_node:
 775                 if (!node)
 776                         break;
 777
 778                 state = rb_entry(node, struct extent_state, rb_node);
 779
 780                 if (state->start > end)
 781                         goto out;
 782
 783                 if (state->state & bits) {
 784                         start = state->start;
 785                         refcount_inc(&state->refs);
 786                         wait_on_state(tree, state);
 787                         free_extent_state(state);
 788                         goto again;
 789                 }
 790                 start = state->end + 1;
 791
 792                 if (start > end)
 793                         break;
 794
 795                 if (!cond_resched_lock(&tree->lock)) {
 796                         node = rb_next(node);
 797                         goto process_node;
 798                 }
 799         }
 800 out:
 801         spin_unlock(&tree->lock);
 802 }
 803
 804 static void set_state_bits(struct extent_io_tree *tree,
 805                            struct extent_state *state,
 806                            unsigned *bits, struct extent_changeset *changeset)
 807 {
 808         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 809
 810         set_state_cb(tree, state, bits);
 811         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 812                 u64 range = state->end - state->start + 1;
 813                 tree->dirty_bytes += range;
 814         }
 815         add_extent_changeset(state, bits_to_set, changeset, 1);
 816         state->state |= bits_to_set;
 817 }
 818
 819 static void cache_state_if_flags(struct extent_state *state,
 820                                  struct extent_state **cached_ptr,
 821                                  unsigned flags)
 822 {
 823         if (cached_ptr && !(*cached_ptr)) {
 824                 if (!flags || (state->state & flags)) {
 825                         *cached_ptr = state;
 826                         refcount_inc(&state->refs);
 827                 }
 828         }
 829 }
 830
 831 static void cache_state(struct extent_state *state,
 832                         struct extent_state **cached_ptr)
 833 {
 834         return cache_state_if_flags(state, cached_ptr,
 835                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 836 }
 837
 838 /*
 839  * set some bits on a range in the tree.  This may require allocations or
 840  * sleeping, so the gfp mask is used to indicate what is allowed.
 841  *
 842  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 843  * part of the range already has the desired bits set.  The start of the
 844  * existing range is returned in failed_start in this case.
 845  *
 846  * [start, end] is inclusive This takes the tree lock.
 847  */
 848
 849 static int __must_check
 850 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 851                  unsigned bits, unsigned exclusive_bits,
 852                  u64 *failed_start, struct extent_state **cached_state,
 853                  gfp_t mask, struct extent_changeset *changeset)
 854 {
 855         struct extent_state *state;
 856         struct extent_state *prealloc = NULL;
 857         struct rb_node *node;
 858         struct rb_node **p;
 859         struct rb_node *parent;
 860         int err = 0;
 861         u64 last_start;
 862         u64 last_end;
 863
 864         btrfs_debug_check_extent_io_range(tree, start, end);
 865
 866         bits |= EXTENT_FIRST_DELALLOC;
 867 again:
 868         if (!prealloc && gfpflags_allow_blocking(mask)) {
 869                 /*
 870                  * Don't care for allocation failure here because we might end
 871                  * up not needing the pre-allocated extent state at all, which
 872                  * is the case if we only have in the tree extent states that
 873                  * cover our input range and don't cover too any other range.
 874                  * If we end up needing a new extent state we allocate it later.
 875                  */
 876                 prealloc = alloc_extent_state(mask);
 877         }
 878
 879         spin_lock(&tree->lock);
 880         if (cached_state && *cached_state) {
 881                 state = *cached_state;
 882                 if (state->start <= start && state->end > start &&
 883                     extent_state_in_tree(state)) {
 884                         node = &state->rb_node;
 885                         goto hit_next;
 886                 }
 887         }
 888         /*
 889          * this search will find all the extents that end after
 890          * our range starts.
 891          */
 892         node = tree_search_for_insert(tree, start, &p, &parent);
 893         if (!node) {
 894                 prealloc = alloc_extent_state_atomic(prealloc);
 895                 BUG_ON(!prealloc);
 896                 err = insert_state(tree, prealloc, start, end,
 897                                    &p, &parent, &bits, changeset);
 898                 if (err)
 899                         extent_io_tree_panic(tree, err);
 900
 901                 cache_state(prealloc, cached_state);
 902                 prealloc = NULL;
 903                 goto out;
 904         }
 905         state = rb_entry(node, struct extent_state, rb_node);
 906 hit_next:
 907         last_start = state->start;
 908         last_end = state->end;
 909
 910         /*
 911          * | ---- desired range ---- |
 912          * | state |
 913          *
 914          * Just lock what we found and keep going
 915          */
 916         if (state->start == start && state->end <= end) {
 917                 if (state->state & exclusive_bits) {
 918                         *failed_start = state->start;
 919                         err = -EEXIST;
 920                         goto out;
 921                 }
 922
 923                 set_state_bits(tree, state, &bits, changeset);
 924                 cache_state(state, cached_state);
 925                 merge_state(tree, state);
 926                 if (last_end == (u64)-1)
 927                         goto out;
 928                 start = last_end + 1;
 929                 state = next_state(state);
 930                 if (start < end && state && state->start == start &&
 931                     !need_resched())
 932                         goto hit_next;
 933                 goto search_again;
 934         }
 935
 936         /*
 937          *     | ---- desired range ---- |
 938          * | state |
 939          *   or
 940          * | ------------- state -------------- |
 941          *
 942          * We need to split the extent we found, and may flip bits on
 943          * second half.
 944          *
 945          * If the extent we found extends past our
 946          * range, we just split and search again.  It'll get split
 947          * again the next time though.
 948          *
 949          * If the extent we found is inside our range, we set the
 950          * desired bit on it.
 951          */
 952         if (state->start < start) {
 953                 if (state->state & exclusive_bits) {
 954                         *failed_start = start;
 955                         err = -EEXIST;
 956                         goto out;
 957                 }
 958
 959                 prealloc = alloc_extent_state_atomic(prealloc);
 960                 BUG_ON(!prealloc);
 961                 err = split_state(tree, state, prealloc, start);
 962                 if (err)
 963                         extent_io_tree_panic(tree, err);
 964
 965                 prealloc = NULL;
 966                 if (err)
 967                         goto out;
 968                 if (state->end <= end) {
 969                         set_state_bits(tree, state, &bits, changeset);
 970                         cache_state(state, cached_state);
 971                         merge_state(tree, state);
 972                         if (last_end == (u64)-1)
 973                                 goto out;
 974                         start = last_end + 1;
 975                         state = next_state(state);
 976                         if (start < end && state && state->start == start &&
 977                             !need_resched())
 978                                 goto hit_next;
 979                 }
 980                 goto search_again;
 981         }
 982         /*
 983          * | ---- desired range ---- |
 984          *     | state | or               | state |
 985          *
 986          * There's a hole, we need to insert something in it and
 987          * ignore the extent we found.
 988          */
 989         if (state->start > start) {
 990                 u64 this_end;
 991                 if (end < last_start)
 992                         this_end = end;
 993                 else
 994                         this_end = last_start - 1;
 995
 996                 prealloc = alloc_extent_state_atomic(prealloc);
 997                 BUG_ON(!prealloc);
 998
 999                 /*
1000                  * Avoid to free 'prealloc' if it can be merged with
1001                  * the later extent.
1002                  */
1003                 err = insert_state(tree, prealloc, start, this_end,
1004                                    NULL, NULL, &bits, changeset);
1005                 if (err)
1006                         extent_io_tree_panic(tree, err);
1007
1008                 cache_state(prealloc, cached_state);
1009                 prealloc = NULL;
1010                 start = this_end + 1;
1011                 goto search_again;
1012         }
1013         /*
1014          * | ---- desired range ---- |
1015          *                        | state |
1016          * We need to split the extent, and set the bit
1017          * on the first half
1018          */
1019         if (state->start <= end && state->end > end) {
1020                 if (state->state & exclusive_bits) {
1021                         *failed_start = start;
1022                         err = -EEXIST;
1023                         goto out;
1024                 }
1025
1026                 prealloc = alloc_extent_state_atomic(prealloc);
1027                 BUG_ON(!prealloc);
1028                 err = split_state(tree, state, prealloc, end + 1);
1029                 if (err)
1030                         extent_io_tree_panic(tree, err);
1031
1032                 set_state_bits(tree, prealloc, &bits, changeset);
1033                 cache_state(prealloc, cached_state);
1034                 merge_state(tree, prealloc);
1035                 prealloc = NULL;
1036                 goto out;
1037         }
1038
1039 search_again:
1040         if (start > end)
1041                 goto out;
1042         spin_unlock(&tree->lock);
1043         if (gfpflags_allow_blocking(mask))
1044                 cond_resched();
1045         goto again;
1046
1047 out:
1048         spin_unlock(&tree->lock);
1049         if (prealloc)
1050                 free_extent_state(prealloc);
1051
1052         return err;
1053
1054 }
1055
1056 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1057                    unsigned bits, u64 * failed_start,
1058                    struct extent_state **cached_state, gfp_t mask)
1059 {
1060         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1061                                 cached_state, mask, NULL);
1062 }
1063
1064
1065 /**
1066  * convert_extent_bit - convert all bits in a given range from one bit to
1067  *                      another
1068  * @tree:       the io tree to search
1069  * @start:      the start offset in bytes
1070  * @end:        the end offset in bytes (inclusive)
1071  * @bits:       the bits to set in this range
1072  * @clear_bits: the bits to clear in this range
1073  * @cached_state:       state that we're going to cache
1074  *
1075  * This will go through and set bits for the given range.  If any states exist
1076  * already in this range they are set with the given bit and cleared of the
1077  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1078  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1079  * boundary bits like LOCK.
1080  *
1081  * All allocations are done with GFP_NOFS.
1082  */
1083 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1084                        unsigned bits, unsigned clear_bits,
1085                        struct extent_state **cached_state)
1086 {
1087         struct extent_state *state;
1088         struct extent_state *prealloc = NULL;
1089         struct rb_node *node;
1090         struct rb_node **p;
1091         struct rb_node *parent;
1092         int err = 0;
1093         u64 last_start;
1094         u64 last_end;
1095         bool first_iteration = true;
1096
1097         btrfs_debug_check_extent_io_range(tree, start, end);
1098
1099 again:
1100         if (!prealloc) {
1101                 /*
1102                  * Best effort, don't worry if extent state allocation fails
1103                  * here for the first iteration. We might have a cached state
1104                  * that matches exactly the target range, in which case no
1105                  * extent state allocations are needed. We'll only know this
1106                  * after locking the tree.
1107                  */
1108                 prealloc = alloc_extent_state(GFP_NOFS);
1109                 if (!prealloc && !first_iteration)
1110                         return -ENOMEM;
1111         }
1112
1113         spin_lock(&tree->lock);
1114         if (cached_state && *cached_state) {
1115                 state = *cached_state;
1116                 if (state->start <= start && state->end > start &&
1117                     extent_state_in_tree(state)) {
1118                         node = &state->rb_node;
1119                         goto hit_next;
1120                 }
1121         }
1122
1123         /*
1124          * this search will find all the extents that end after
1125          * our range starts.
1126          */
1127         node = tree_search_for_insert(tree, start, &p, &parent);
1128         if (!node) {
1129                 prealloc = alloc_extent_state_atomic(prealloc);
1130                 if (!prealloc) {
1131                         err = -ENOMEM;
1132                         goto out;
1133                 }
1134                 err = insert_state(tree, prealloc, start, end,
1135                                    &p, &parent, &bits, NULL);
1136                 if (err)
1137                         extent_io_tree_panic(tree, err);
1138                 cache_state(prealloc, cached_state);
1139                 prealloc = NULL;
1140                 goto out;
1141         }
1142         state = rb_entry(node, struct extent_state, rb_node);
1143 hit_next:
1144         last_start = state->start;
1145         last_end = state->end;
1146
1147         /*
1148          * | ---- desired range ---- |
1149          * | state |
1150          *
1151          * Just lock what we found and keep going
1152          */
1153         if (state->start == start && state->end <= end) {
1154                 set_state_bits(tree, state, &bits, NULL);
1155                 cache_state(state, cached_state);
1156                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1157                 if (last_end == (u64)-1)
1158                         goto out;
1159                 start = last_end + 1;
1160                 if (start < end && state && state->start == start &&
1161                     !need_resched())
1162                         goto hit_next;
1163                 goto search_again;
1164         }
1165
1166         /*
1167          *     | ---- desired range ---- |
1168          * | state |
1169          *   or
1170          * | ------------- state -------------- |
1171          *
1172          * We need to split the extent we found, and may flip bits on
1173          * second half.
1174          *
1175          * If the extent we found extends past our
1176          * range, we just split and search again.  It'll get split
1177          * again the next time though.
1178          *
1179          * If the extent we found is inside our range, we set the
1180          * desired bit on it.
1181          */
1182         if (state->start < start) {
1183                 prealloc = alloc_extent_state_atomic(prealloc);
1184                 if (!prealloc) {
1185                         err = -ENOMEM;
1186                         goto out;
1187                 }
1188                 err = split_state(tree, state, prealloc, start);
1189                 if (err)
1190                         extent_io_tree_panic(tree, err);
1191                 prealloc = NULL;
1192                 if (err)
1193                         goto out;
1194                 if (state->end <= end) {
1195                         set_state_bits(tree, state, &bits, NULL);
1196                         cache_state(state, cached_state);
1197                         state = clear_state_bit(tree, state, &clear_bits, 0,
1198                                                 NULL);
1199                         if (last_end == (u64)-1)
1200                                 goto out;
1201                         start = last_end + 1;
1202                         if (start < end && state && state->start == start &&
1203                             !need_resched())
1204                                 goto hit_next;
1205                 }
1206                 goto search_again;
1207         }
1208         /*
1209          * | ---- desired range ---- |
1210          *     | state | or               | state |
1211          *
1212          * There's a hole, we need to insert something in it and
1213          * ignore the extent we found.
1214          */
1215         if (state->start > start) {
1216                 u64 this_end;
1217                 if (end < last_start)
1218                         this_end = end;
1219                 else
1220                         this_end = last_start - 1;
1221
1222                 prealloc = alloc_extent_state_atomic(prealloc);
1223                 if (!prealloc) {
1224                         err = -ENOMEM;
1225                         goto out;
1226                 }
1227
1228                 /*
1229                  * Avoid to free 'prealloc' if it can be merged with
1230                  * the later extent.
1231                  */
1232                 err = insert_state(tree, prealloc, start, this_end,
1233                                    NULL, NULL, &bits, NULL);
1234                 if (err)
1235                         extent_io_tree_panic(tree, err);
1236                 cache_state(prealloc, cached_state);
1237                 prealloc = NULL;
1238                 start = this_end + 1;
1239                 goto search_again;
1240         }
1241         /*
1242          * | ---- desired range ---- |
1243          *                        | state |
1244          * We need to split the extent, and set the bit
1245          * on the first half
1246          */
1247         if (state->start <= end && state->end > end) {
1248                 prealloc = alloc_extent_state_atomic(prealloc);
1249                 if (!prealloc) {
1250                         err = -ENOMEM;
1251                         goto out;
1252                 }
1253
1254                 err = split_state(tree, state, prealloc, end + 1);
1255                 if (err)
1256                         extent_io_tree_panic(tree, err);
1257
1258                 set_state_bits(tree, prealloc, &bits, NULL);
1259                 cache_state(prealloc, cached_state);
1260                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1261                 prealloc = NULL;
1262                 goto out;
1263         }
1264
1265 search_again:
1266         if (start > end)
1267                 goto out;
1268         spin_unlock(&tree->lock);
1269         cond_resched();
1270         first_iteration = false;
1271         goto again;
1272
1273 out:
1274         spin_unlock(&tree->lock);
1275         if (prealloc)
1276                 free_extent_state(prealloc);
1277
1278         return err;
1279 }
1280
1281 /* wrappers around set/clear extent bit */
1282 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1283                            unsigned bits, struct extent_changeset *changeset)
1284 {
1285         /*
1286          * We don't support EXTENT_LOCKED yet, as current changeset will
1287          * record any bits changed, so for EXTENT_LOCKED case, it will
1288          * either fail with -EEXIST or changeset will record the whole
1289          * range.
1290          */
1291         BUG_ON(bits & EXTENT_LOCKED);
1292
1293         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1294                                 changeset);
1295 }
1296
1297 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1298                      unsigned bits, int wake, int delete,
1299                      struct extent_state **cached, gfp_t mask)
1300 {
1301         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1302                                   cached, mask, NULL);
1303 }
1304
1305 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1306                 unsigned bits, struct extent_changeset *changeset)
1307 {
1308         /*
1309          * Don't support EXTENT_LOCKED case, same reason as
1310          * set_record_extent_bits().
1311          */
1312         BUG_ON(bits & EXTENT_LOCKED);
1313
1314         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1315                                   changeset);
1316 }
1317
1318 /*
1319  * either insert or lock state struct between start and end use mask to tell
1320  * us if waiting is desired.
1321  */
1322 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1323                      struct extent_state **cached_state)
1324 {
1325         int err;
1326         u64 failed_start;
1327
1328         while (1) {
1329                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1330                                        EXTENT_LOCKED, &failed_start,
1331                                        cached_state, GFP_NOFS, NULL);
1332                 if (err == -EEXIST) {
1333                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1334                         start = failed_start;
1335                 } else
1336                         break;
1337                 WARN_ON(start > end);
1338         }
1339         return err;
1340 }
1341
1342 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1343 {
1344         int err;
1345         u64 failed_start;
1346
1347         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1348                                &failed_start, NULL, GFP_NOFS, NULL);
1349         if (err == -EEXIST) {
1350                 if (failed_start > start)
1351                         clear_extent_bit(tree, start, failed_start - 1,
1352                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1353                 return 0;
1354         }
1355         return 1;
1356 }
1357
1358 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1359 {
1360         unsigned long index = start >> PAGE_SHIFT;
1361         unsigned long end_index = end >> PAGE_SHIFT;
1362         struct page *page;
1363
1364         while (index <= end_index) {
1365                 page = find_get_page(inode->i_mapping, index);
1366                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1367                 clear_page_dirty_for_io(page);
1368                 put_page(page);
1369                 index++;
1370         }
1371 }
1372
1373 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1374 {
1375         unsigned long index = start >> PAGE_SHIFT;
1376         unsigned long end_index = end >> PAGE_SHIFT;
1377         struct page *page;
1378
1379         while (index <= end_index) {
1380                 page = find_get_page(inode->i_mapping, index);
1381                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1382                 __set_page_dirty_nobuffers(page);
1383                 account_page_redirty(page);
1384                 put_page(page);
1385                 index++;
1386         }
1387 }
1388
1389 /*
1390  * helper function to set both pages and extents in the tree writeback
1391  */
1392 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1393 {
1394         tree->ops->set_range_writeback(tree->private_data, start, end);
1395 }
1396
1397 /* find the first state struct with 'bits' set after 'start', and
1398  * return it.  tree->lock must be held.  NULL will returned if
1399  * nothing was found after 'start'
1400  */
1401 static struct extent_state *
1402 find_first_extent_bit_state(struct extent_io_tree *tree,
1403                             u64 start, unsigned bits)
1404 {
1405         struct rb_node *node;
1406         struct extent_state *state;
1407
1408         /*
1409          * this search will find all the extents that end after
1410          * our range starts.
1411          */
1412         node = tree_search(tree, start);
1413         if (!node)
1414                 goto out;
1415
1416         while (1) {
1417                 state = rb_entry(node, struct extent_state, rb_node);
1418                 if (state->end >= start && (state->state & bits))
1419                         return state;
1420
1421                 node = rb_next(node);
1422                 if (!node)
1423                         break;
1424         }
1425 out:
1426         return NULL;
1427 }
1428
1429 /*
1430  * find the first offset in the io tree with 'bits' set. zero is
1431  * returned if we find something, and *start_ret and *end_ret are
1432  * set to reflect the state struct that was found.
1433  *
1434  * If nothing was found, 1 is returned. If found something, return 0.
1435  */
1436 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1437                           u64 *start_ret, u64 *end_ret, unsigned bits,
1438                           struct extent_state **cached_state)
1439 {
1440         struct extent_state *state;
1441         struct rb_node *n;
1442         int ret = 1;
1443
1444         spin_lock(&tree->lock);
1445         if (cached_state && *cached_state) {
1446                 state = *cached_state;
1447                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1448                         n = rb_next(&state->rb_node);
1449                         while (n) {
1450                                 state = rb_entry(n, struct extent_state,
1451                                                  rb_node);
1452                                 if (state->state & bits)
1453                                         goto got_it;
1454                                 n = rb_next(n);
1455                         }
1456                         free_extent_state(*cached_state);
1457                         *cached_state = NULL;
1458                         goto out;
1459                 }
1460                 free_extent_state(*cached_state);
1461                 *cached_state = NULL;
1462         }
1463
1464         state = find_first_extent_bit_state(tree, start, bits);
1465 got_it:
1466         if (state) {
1467                 cache_state_if_flags(state, cached_state, 0);
1468                 *start_ret = state->start;
1469                 *end_ret = state->end;
1470                 ret = 0;
1471         }
1472 out:
1473         spin_unlock(&tree->lock);
1474         return ret;
1475 }
1476
1477 /*
1478  * find a contiguous range of bytes in the file marked as delalloc, not
1479  * more than 'max_bytes'.  start and end are used to return the range,
1480  *
1481  * 1 is returned if we find something, 0 if nothing was in the tree
1482  */
1483 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1484                                         u64 *start, u64 *end, u64 max_bytes,
1485                                         struct extent_state **cached_state)
1486 {
1487         struct rb_node *node;
1488         struct extent_state *state;
1489         u64 cur_start = *start;
1490         u64 found = 0;
1491         u64 total_bytes = 0;
1492
1493         spin_lock(&tree->lock);
1494
1495         /*
1496          * this search will find all the extents that end after
1497          * our range starts.
1498          */
1499         node = tree_search(tree, cur_start);
1500         if (!node) {
1501                 if (!found)
1502                         *end = (u64)-1;
1503                 goto out;
1504         }
1505
1506         while (1) {
1507                 state = rb_entry(node, struct extent_state, rb_node);
1508                 if (found && (state->start != cur_start ||
1509                               (state->state & EXTENT_BOUNDARY))) {
1510                         goto out;
1511                 }
1512                 if (!(state->state & EXTENT_DELALLOC)) {
1513                         if (!found)
1514                                 *end = state->end;
1515                         goto out;
1516                 }
1517                 if (!found) {
1518                         *start = state->start;
1519                         *cached_state = state;
1520                         refcount_inc(&state->refs);
1521                 }
1522                 found++;
1523                 *end = state->end;
1524                 cur_start = state->end + 1;
1525                 node = rb_next(node);
1526                 total_bytes += state->end - state->start + 1;
1527                 if (total_bytes >= max_bytes)
1528                         break;
1529                 if (!node)
1530                         break;
1531         }
1532 out:
1533         spin_unlock(&tree->lock);
1534         return found;
1535 }
1536
1537 static int __process_pages_contig(struct address_space *mapping,
1538                                   struct page *locked_page,
1539                                   pgoff_t start_index, pgoff_t end_index,
1540                                   unsigned long page_ops, pgoff_t *index_ret);
1541
1542 static noinline void __unlock_for_delalloc(struct inode *inode,
1543                                            struct page *locked_page,
1544                                            u64 start, u64 end)
1545 {
1546         unsigned long index = start >> PAGE_SHIFT;
1547         unsigned long end_index = end >> PAGE_SHIFT;
1548
1549         ASSERT(locked_page);
1550         if (index == locked_page->index && end_index == index)
1551                 return;
1552
1553         __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1554                                PAGE_UNLOCK, NULL);
1555 }
1556
1557 static noinline int lock_delalloc_pages(struct inode *inode,
1558                                         struct page *locked_page,
1559                                         u64 delalloc_start,
1560                                         u64 delalloc_end)
1561 {
1562         unsigned long index = delalloc_start >> PAGE_SHIFT;
1563         unsigned long index_ret = index;
1564         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1565         int ret;
1566
1567         ASSERT(locked_page);
1568         if (index == locked_page->index && index == end_index)
1569                 return 0;
1570
1571         ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1572                                      end_index, PAGE_LOCK, &index_ret);
1573         if (ret == -EAGAIN)
1574                 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1575                                       (u64)index_ret << PAGE_SHIFT);
1576         return ret;
1577 }
1578
1579 /*
1580  * find a contiguous range of bytes in the file marked as delalloc, not
1581  * more than 'max_bytes'.  start and end are used to return the range,
1582  *
1583  * 1 is returned if we find something, 0 if nothing was in the tree
1584  */
1585 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1586                                     struct extent_io_tree *tree,
1587                                     struct page *locked_page, u64 *start,
1588                                     u64 *end, u64 max_bytes)
1589 {
1590         u64 delalloc_start;
1591         u64 delalloc_end;
1592         u64 found;
1593         struct extent_state *cached_state = NULL;
1594         int ret;
1595         int loops = 0;
1596
1597 again:
1598         /* step one, find a bunch of delalloc bytes starting at start */
1599         delalloc_start = *start;
1600         delalloc_end = 0;
1601         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1602                                     max_bytes, &cached_state);
1603         if (!found || delalloc_end <= *start) {
1604                 *start = delalloc_start;
1605                 *end = delalloc_end;
1606                 free_extent_state(cached_state);
1607                 return 0;
1608         }
1609
1610         /*
1611          * start comes from the offset of locked_page.  We have to lock
1612          * pages in order, so we can't process delalloc bytes before
1613          * locked_page
1614          */
1615         if (delalloc_start < *start)
1616                 delalloc_start = *start;
1617
1618         /*
1619          * make sure to limit the number of pages we try to lock down
1620          */
1621         if (delalloc_end + 1 - delalloc_start > max_bytes)
1622                 delalloc_end = delalloc_start + max_bytes - 1;
1623
1624         /* step two, lock all the pages after the page that has start */
1625         ret = lock_delalloc_pages(inode, locked_page,
1626                                   delalloc_start, delalloc_end);
1627         if (ret == -EAGAIN) {
1628                 /* some of the pages are gone, lets avoid looping by
1629                  * shortening the size of the delalloc range we're searching
1630                  */
1631                 free_extent_state(cached_state);
1632                 cached_state = NULL;
1633                 if (!loops) {
1634                         max_bytes = PAGE_SIZE;
1635                         loops = 1;
1636                         goto again;
1637                 } else {
1638                         found = 0;
1639                         goto out_failed;
1640                 }
1641         }
1642         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1643
1644         /* step three, lock the state bits for the whole range */
1645         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1646
1647         /* then test to make sure it is all still delalloc */
1648         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1649                              EXTENT_DELALLOC, 1, cached_state);
1650         if (!ret) {
1651                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1652                                      &cached_state, GFP_NOFS);
1653                 __unlock_for_delalloc(inode, locked_page,
1654                               delalloc_start, delalloc_end);
1655                 cond_resched();
1656                 goto again;
1657         }
1658         free_extent_state(cached_state);
1659         *start = delalloc_start;
1660         *end = delalloc_end;
1661 out_failed:
1662         return found;
1663 }
1664
1665 static int __process_pages_contig(struct address_space *mapping,
1666                                   struct page *locked_page,
1667                                   pgoff_t start_index, pgoff_t end_index,
1668                                   unsigned long page_ops, pgoff_t *index_ret)
1669 {
1670         unsigned long nr_pages = end_index - start_index + 1;
1671         unsigned long pages_locked = 0;
1672         pgoff_t index = start_index;
1673         struct page *pages[16];
1674         unsigned ret;
1675         int err = 0;
1676         int i;
1677
1678         if (page_ops & PAGE_LOCK) {
1679                 ASSERT(page_ops == PAGE_LOCK);
1680                 ASSERT(index_ret && *index_ret == start_index);
1681         }
1682
1683         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1684                 mapping_set_error(mapping, -EIO);
1685
1686         while (nr_pages > 0) {
1687                 ret = find_get_pages_contig(mapping, index,
1688                                      min_t(unsigned long,
1689                                      nr_pages, ARRAY_SIZE(pages)), pages);
1690                 if (ret == 0) {
1691                         /*
1692                          * Only if we're going to lock these pages,
1693                          * can we find nothing at @index.
1694                          */
1695                         ASSERT(page_ops & PAGE_LOCK);
1696                         err = -EAGAIN;
1697                         goto out;
1698                 }
1699
1700                 for (i = 0; i < ret; i++) {
1701                         if (page_ops & PAGE_SET_PRIVATE2)
1702                                 SetPagePrivate2(pages[i]);
1703
1704                         if (pages[i] == locked_page) {
1705                                 put_page(pages[i]);
1706                                 pages_locked++;
1707                                 continue;
1708                         }
1709                         if (page_ops & PAGE_CLEAR_DIRTY)
1710                                 clear_page_dirty_for_io(pages[i]);
1711                         if (page_ops & PAGE_SET_WRITEBACK)
1712                                 set_page_writeback(pages[i]);
1713                         if (page_ops & PAGE_SET_ERROR)
1714                                 SetPageError(pages[i]);
1715                         if (page_ops & PAGE_END_WRITEBACK)
1716                                 end_page_writeback(pages[i]);
1717                         if (page_ops & PAGE_UNLOCK)
1718                                 unlock_page(pages[i]);
1719                         if (page_ops & PAGE_LOCK) {
1720                                 lock_page(pages[i]);
1721                                 if (!PageDirty(pages[i]) ||
1722                                     pages[i]->mapping != mapping) {
1723                                         unlock_page(pages[i]);
1724                                         put_page(pages[i]);
1725                                         err = -EAGAIN;
1726                                         goto out;
1727                                 }
1728                         }
1729                         put_page(pages[i]);
1730                         pages_locked++;
1731                 }
1732                 nr_pages -= ret;
1733                 index += ret;
1734                 cond_resched();
1735         }
1736 out:
1737         if (err && index_ret)
1738                 *index_ret = start_index + pages_locked - 1;
1739         return err;
1740 }
1741
1742 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1743                                  u64 delalloc_end, struct page *locked_page,
1744                                  unsigned clear_bits,
1745                                  unsigned long page_ops)
1746 {
1747         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1748                          NULL, GFP_NOFS);
1749
1750         __process_pages_contig(inode->i_mapping, locked_page,
1751                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1752                                page_ops, NULL);
1753 }
1754
1755 /*
1756  * count the number of bytes in the tree that have a given bit(s)
1757  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1758  * cached.  The total number found is returned.
1759  */
1760 u64 count_range_bits(struct extent_io_tree *tree,
1761                      u64 *start, u64 search_end, u64 max_bytes,
1762                      unsigned bits, int contig)
1763 {
1764         struct rb_node *node;
1765         struct extent_state *state;
1766         u64 cur_start = *start;
1767         u64 total_bytes = 0;
1768         u64 last = 0;
1769         int found = 0;
1770
1771         if (WARN_ON(search_end <= cur_start))
1772                 return 0;
1773
1774         spin_lock(&tree->lock);
1775         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1776                 total_bytes = tree->dirty_bytes;
1777                 goto out;
1778         }
1779         /*
1780          * this search will find all the extents that end after
1781          * our range starts.
1782          */
1783         node = tree_search(tree, cur_start);
1784         if (!node)
1785                 goto out;
1786
1787         while (1) {
1788                 state = rb_entry(node, struct extent_state, rb_node);
1789                 if (state->start > search_end)
1790                         break;
1791                 if (contig && found && state->start > last + 1)
1792                         break;
1793                 if (state->end >= cur_start && (state->state & bits) == bits) {
1794                         total_bytes += min(search_end, state->end) + 1 -
1795                                        max(cur_start, state->start);
1796                         if (total_bytes >= max_bytes)
1797                                 break;
1798                         if (!found) {
1799                                 *start = max(cur_start, state->start);
1800                                 found = 1;
1801                         }
1802                         last = state->end;
1803                 } else if (contig && found) {
1804                         break;
1805                 }
1806                 node = rb_next(node);
1807                 if (!node)
1808                         break;
1809         }
1810 out:
1811         spin_unlock(&tree->lock);
1812         return total_bytes;
1813 }
1814
1815 /*
1816  * set the private field for a given byte offset in the tree.  If there isn't
1817  * an extent_state there already, this does nothing.
1818  */
1819 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1820                 struct io_failure_record *failrec)
1821 {
1822         struct rb_node *node;
1823         struct extent_state *state;
1824         int ret = 0;
1825
1826         spin_lock(&tree->lock);
1827         /*
1828          * this search will find all the extents that end after
1829          * our range starts.
1830          */
1831         node = tree_search(tree, start);
1832         if (!node) {
1833                 ret = -ENOENT;
1834                 goto out;
1835         }
1836         state = rb_entry(node, struct extent_state, rb_node);
1837         if (state->start != start) {
1838                 ret = -ENOENT;
1839                 goto out;
1840         }
1841         state->failrec = failrec;
1842 out:
1843         spin_unlock(&tree->lock);
1844         return ret;
1845 }
1846
1847 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1848                 struct io_failure_record **failrec)
1849 {
1850         struct rb_node *node;
1851         struct extent_state *state;
1852         int ret = 0;
1853
1854         spin_lock(&tree->lock);
1855         /*
1856          * this search will find all the extents that end after
1857          * our range starts.
1858          */
1859         node = tree_search(tree, start);
1860         if (!node) {
1861                 ret = -ENOENT;
1862                 goto out;
1863         }
1864         state = rb_entry(node, struct extent_state, rb_node);
1865         if (state->start != start) {
1866                 ret = -ENOENT;
1867                 goto out;
1868         }
1869         *failrec = state->failrec;
1870 out:
1871         spin_unlock(&tree->lock);
1872         return ret;
1873 }
1874
1875 /*
1876  * searches a range in the state tree for a given mask.
1877  * If 'filled' == 1, this returns 1 only if every extent in the tree
1878  * has the bits set.  Otherwise, 1 is returned if any bit in the
1879  * range is found set.
1880  */
1881 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1882                    unsigned bits, int filled, struct extent_state *cached)
1883 {
1884         struct extent_state *state = NULL;
1885         struct rb_node *node;
1886         int bitset = 0;
1887
1888         spin_lock(&tree->lock);
1889         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1890             cached->end > start)
1891                 node = &cached->rb_node;
1892         else
1893                 node = tree_search(tree, start);
1894         while (node && start <= end) {
1895                 state = rb_entry(node, struct extent_state, rb_node);
1896
1897                 if (filled && state->start > start) {
1898                         bitset = 0;
1899                         break;
1900                 }
1901
1902                 if (state->start > end)
1903                         break;
1904
1905                 if (state->state & bits) {
1906                         bitset = 1;
1907                         if (!filled)
1908                                 break;
1909                 } else if (filled) {
1910                         bitset = 0;
1911                         break;
1912                 }
1913
1914                 if (state->end == (u64)-1)
1915                         break;
1916
1917                 start = state->end + 1;
1918                 if (start > end)
1919                         break;
1920                 node = rb_next(node);
1921                 if (!node) {
1922                         if (filled)
1923                                 bitset = 0;
1924                         break;
1925                 }
1926         }
1927         spin_unlock(&tree->lock);
1928         return bitset;
1929 }
1930
1931 /*
1932  * helper function to set a given page up to date if all the
1933  * extents in the tree for that page are up to date
1934  */
1935 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1936 {
1937         u64 start = page_offset(page);
1938         u64 end = start + PAGE_SIZE - 1;
1939         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1940                 SetPageUptodate(page);
1941 }
1942
1943 int free_io_failure(struct extent_io_tree *failure_tree,
1944                     struct extent_io_tree *io_tree,
1945                     struct io_failure_record *rec)
1946 {
1947         int ret;
1948         int err = 0;
1949
1950         set_state_failrec(failure_tree, rec->start, NULL);
1951         ret = clear_extent_bits(failure_tree, rec->start,
1952                                 rec->start + rec->len - 1,
1953                                 EXTENT_LOCKED | EXTENT_DIRTY);
1954         if (ret)
1955                 err = ret;
1956
1957         ret = clear_extent_bits(io_tree, rec->start,
1958                                 rec->start + rec->len - 1,
1959                                 EXTENT_DAMAGED);
1960         if (ret && !err)
1961                 err = ret;
1962
1963         kfree(rec);
1964         return err;
1965 }
1966
1967 /*
1968  * this bypasses the standard btrfs submit functions deliberately, as
1969  * the standard behavior is to write all copies in a raid setup. here we only
1970  * want to write the one bad copy. so we do the mapping for ourselves and issue
1971  * submit_bio directly.
1972  * to avoid any synchronization issues, wait for the data after writing, which
1973  * actually prevents the read that triggered the error from finishing.
1974  * currently, there can be no more than two copies of every data bit. thus,
1975  * exactly one rewrite is required.
1976  */
1977 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
1978                       u64 length, u64 logical, struct page *page,
1979                       unsigned int pg_offset, int mirror_num)
1980 {
1981         struct bio *bio;
1982         struct btrfs_device *dev;
1983         u64 map_length = 0;
1984         u64 sector;
1985         struct btrfs_bio *bbio = NULL;
1986         int ret;
1987
1988         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
1989         BUG_ON(!mirror_num);
1990
1991         bio = btrfs_io_bio_alloc(1);
1992         bio->bi_iter.bi_size = 0;
1993         map_length = length;
1994
1995         /*
1996          * Avoid races with device replace and make sure our bbio has devices
1997          * associated to its stripes that don't go away while we are doing the
1998          * read repair operation.
1999          */
2000         btrfs_bio_counter_inc_blocked(fs_info);
2001         if (btrfs_is_parity_mirror(fs_info, logical, length, mirror_num)) {
2002                 /*
2003                  * Note that we don't use BTRFS_MAP_WRITE because it's supposed
2004                  * to update all raid stripes, but here we just want to correct
2005                  * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
2006                  * stripe's dev and sector.
2007                  */
2008                 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
2009                                       &map_length, &bbio, 0);
2010                 if (ret) {
2011                         btrfs_bio_counter_dec(fs_info);
2012                         bio_put(bio);
2013                         return -EIO;
2014                 }
2015                 ASSERT(bbio->mirror_num == 1);
2016         } else {
2017                 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2018                                       &map_length, &bbio, mirror_num);
2019                 if (ret) {
2020                         btrfs_bio_counter_dec(fs_info);
2021                         bio_put(bio);
2022                         return -EIO;
2023                 }
2024                 BUG_ON(mirror_num != bbio->mirror_num);
2025         }
2026
2027         sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2028         bio->bi_iter.bi_sector = sector;
2029         dev = bbio->stripes[bbio->mirror_num - 1].dev;
2030         btrfs_put_bbio(bbio);
2031         if (!dev || !dev->bdev || !dev->writeable) {
2032                 btrfs_bio_counter_dec(fs_info);
2033                 bio_put(bio);
2034                 return -EIO;
2035         }
2036         bio->bi_bdev = dev->bdev;
2037         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2038         bio_add_page(bio, page, length, pg_offset);
2039
2040         if (btrfsic_submit_bio_wait(bio)) {
2041                 /* try to remap that extent elsewhere? */
2042                 btrfs_bio_counter_dec(fs_info);
2043                 bio_put(bio);
2044                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2045                 return -EIO;
2046         }
2047
2048         btrfs_info_rl_in_rcu(fs_info,
2049                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2050                                   ino, start,
2051                                   rcu_str_deref(dev->name), sector);
2052         btrfs_bio_counter_dec(fs_info);
2053         bio_put(bio);
2054         return 0;
2055 }
2056
2057 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2058                          struct extent_buffer *eb, int mirror_num)
2059 {
2060         u64 start = eb->start;
2061         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2062         int ret = 0;
2063
2064         if (fs_info->sb->s_flags & MS_RDONLY)
2065                 return -EROFS;
2066
2067         for (i = 0; i < num_pages; i++) {
2068                 struct page *p = eb->pages[i];
2069
2070                 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
2071                                         start - page_offset(p), mirror_num);
2072                 if (ret)
2073                         break;
2074                 start += PAGE_SIZE;
2075         }
2076
2077         return ret;
2078 }
2079
2080 /*
2081  * each time an IO finishes, we do a fast check in the IO failure tree
2082  * to see if we need to process or clean up an io_failure_record
2083  */
2084 int clean_io_failure(struct btrfs_fs_info *fs_info,
2085                      struct extent_io_tree *failure_tree,
2086                      struct extent_io_tree *io_tree, u64 start,
2087                      struct page *page, u64 ino, unsigned int pg_offset)
2088 {
2089         u64 private;
2090         struct io_failure_record *failrec;
2091         struct extent_state *state;
2092         int num_copies;
2093         int ret;
2094
2095         private = 0;
2096         ret = count_range_bits(failure_tree, &private, (u64)-1, 1,
2097                                EXTENT_DIRTY, 0);
2098         if (!ret)
2099                 return 0;
2100
2101         ret = get_state_failrec(failure_tree, start, &failrec);
2102         if (ret)
2103                 return 0;
2104
2105         BUG_ON(!failrec->this_mirror);
2106
2107         if (failrec->in_validation) {
2108                 /* there was no real error, just free the record */
2109                 btrfs_debug(fs_info,
2110                         "clean_io_failure: freeing dummy error at %llu",
2111                         failrec->start);
2112                 goto out;
2113         }
2114         if (fs_info->sb->s_flags & MS_RDONLY)
2115                 goto out;
2116
2117         spin_lock(&io_tree->lock);
2118         state = find_first_extent_bit_state(io_tree,
2119                                             failrec->start,
2120                                             EXTENT_LOCKED);
2121         spin_unlock(&io_tree->lock);
2122
2123         if (state && state->start <= failrec->start &&
2124             state->end >= failrec->start + failrec->len - 1) {
2125                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2126                                               failrec->len);
2127                 if (num_copies > 1)  {
2128                         repair_io_failure(fs_info, ino, start, failrec->len,
2129                                           failrec->logical, page, pg_offset,
2130                                           failrec->failed_mirror);
2131                 }
2132         }
2133
2134 out:
2135         free_io_failure(failure_tree, io_tree, failrec);
2136
2137         return 0;
2138 }
2139
2140 /*
2141  * Can be called when
2142  * - hold extent lock
2143  * - under ordered extent
2144  * - the inode is freeing
2145  */
2146 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2147 {
2148         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2149         struct io_failure_record *failrec;
2150         struct extent_state *state, *next;
2151
2152         if (RB_EMPTY_ROOT(&failure_tree->state))
2153                 return;
2154
2155         spin_lock(&failure_tree->lock);
2156         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2157         while (state) {
2158                 if (state->start > end)
2159                         break;
2160
2161                 ASSERT(state->end <= end);
2162
2163                 next = next_state(state);
2164
2165                 failrec = state->failrec;
2166                 free_extent_state(state);
2167                 kfree(failrec);
2168
2169                 state = next;
2170         }
2171         spin_unlock(&failure_tree->lock);
2172 }
2173
2174 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2175                 struct io_failure_record **failrec_ret)
2176 {
2177         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2178         struct io_failure_record *failrec;
2179         struct extent_map *em;
2180         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2181         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2182         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2183         int ret;
2184         u64 logical;
2185
2186         ret = get_state_failrec(failure_tree, start, &failrec);
2187         if (ret) {
2188                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2189                 if (!failrec)
2190                         return -ENOMEM;
2191
2192                 failrec->start = start;
2193                 failrec->len = end - start + 1;
2194                 failrec->this_mirror = 0;
2195                 failrec->bio_flags = 0;
2196                 failrec->in_validation = 0;
2197
2198                 read_lock(&em_tree->lock);
2199                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2200                 if (!em) {
2201                         read_unlock(&em_tree->lock);
2202                         kfree(failrec);
2203                         return -EIO;
2204                 }
2205
2206                 if (em->start > start || em->start + em->len <= start) {
2207                         free_extent_map(em);
2208                         em = NULL;
2209                 }
2210                 read_unlock(&em_tree->lock);
2211                 if (!em) {
2212                         kfree(failrec);
2213                         return -EIO;
2214                 }
2215
2216                 logical = start - em->start;
2217                 logical = em->block_start + logical;
2218                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2219                         logical = em->block_start;
2220                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2221                         extent_set_compress_type(&failrec->bio_flags,
2222                                                  em->compress_type);
2223                 }
2224
2225                 btrfs_debug(fs_info,
2226                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2227                         logical, start, failrec->len);
2228
2229                 failrec->logical = logical;
2230                 free_extent_map(em);
2231
2232                 /* set the bits in the private failure tree */
2233                 ret = set_extent_bits(failure_tree, start, end,
2234                                         EXTENT_LOCKED | EXTENT_DIRTY);
2235                 if (ret >= 0)
2236                         ret = set_state_failrec(failure_tree, start, failrec);
2237                 /* set the bits in the inode's tree */
2238                 if (ret >= 0)
2239                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2240                 if (ret < 0) {
2241                         kfree(failrec);
2242                         return ret;
2243                 }
2244         } else {
2245                 btrfs_debug(fs_info,
2246                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2247                         failrec->logical, failrec->start, failrec->len,
2248                         failrec->in_validation);
2249                 /*
2250                  * when data can be on disk more than twice, add to failrec here
2251                  * (e.g. with a list for failed_mirror) to make
2252                  * clean_io_failure() clean all those errors at once.
2253                  */
2254         }
2255
2256         *failrec_ret = failrec;
2257
2258         return 0;
2259 }
2260
2261 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2262                            struct io_failure_record *failrec, int failed_mirror)
2263 {
2264         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2265         int num_copies;
2266
2267         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2268         if (num_copies == 1) {
2269                 /*
2270                  * we only have a single copy of the data, so don't bother with
2271                  * all the retry and error correction code that follows. no
2272                  * matter what the error is, it is very likely to persist.
2273                  */
2274                 btrfs_debug(fs_info,
2275                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2276                         num_copies, failrec->this_mirror, failed_mirror);
2277                 return 0;
2278         }
2279
2280         /*
2281          * there are two premises:
2282          *      a) deliver good data to the caller
2283          *      b) correct the bad sectors on disk
2284          */
2285         if (failed_bio->bi_vcnt > 1) {
2286                 /*
2287                  * to fulfill b), we need to know the exact failing sectors, as
2288                  * we don't want to rewrite any more than the failed ones. thus,
2289                  * we need separate read requests for the failed bio
2290                  *
2291                  * if the following BUG_ON triggers, our validation request got
2292                  * merged. we need separate requests for our algorithm to work.
2293                  */
2294                 BUG_ON(failrec->in_validation);
2295                 failrec->in_validation = 1;
2296                 failrec->this_mirror = failed_mirror;
2297         } else {
2298                 /*
2299                  * we're ready to fulfill a) and b) alongside. get a good copy
2300                  * of the failed sector and if we succeed, we have setup
2301                  * everything for repair_io_failure to do the rest for us.
2302                  */
2303                 if (failrec->in_validation) {
2304                         BUG_ON(failrec->this_mirror != failed_mirror);
2305                         failrec->in_validation = 0;
2306                         failrec->this_mirror = 0;
2307                 }
2308                 failrec->failed_mirror = failed_mirror;
2309                 failrec->this_mirror++;
2310                 if (failrec->this_mirror == failed_mirror)
2311                         failrec->this_mirror++;
2312         }
2313
2314         if (failrec->this_mirror > num_copies) {
2315                 btrfs_debug(fs_info,
2316                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2317                         num_copies, failrec->this_mirror, failed_mirror);
2318                 return 0;
2319         }
2320
2321         return 1;
2322 }
2323
2324
2325 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2326                                     struct io_failure_record *failrec,
2327                                     struct page *page, int pg_offset, int icsum,
2328                                     bio_end_io_t *endio_func, void *data)
2329 {
2330         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2331         struct bio *bio;
2332         struct btrfs_io_bio *btrfs_failed_bio;
2333         struct btrfs_io_bio *btrfs_bio;
2334
2335         bio = btrfs_io_bio_alloc(1);
2336         bio->bi_end_io = endio_func;
2337         bio->bi_iter.bi_sector = failrec->logical >> 9;
2338         bio->bi_bdev = fs_info->fs_devices->latest_bdev;
2339         bio->bi_iter.bi_size = 0;
2340         bio->bi_private = data;
2341
2342         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2343         if (btrfs_failed_bio->csum) {
2344                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2345
2346                 btrfs_bio = btrfs_io_bio(bio);
2347                 btrfs_bio->csum = btrfs_bio->csum_inline;
2348                 icsum *= csum_size;
2349                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2350                        csum_size);
2351         }
2352
2353         bio_add_page(bio, page, failrec->len, pg_offset);
2354
2355         return bio;
2356 }
2357
2358 /*
2359  * this is a generic handler for readpage errors (default
2360  * readpage_io_failed_hook). if other copies exist, read those and write back
2361  * good data to the failed position. does not investigate in remapping the
2362  * failed extent elsewhere, hoping the device will be smart enough to do this as
2363  * needed
2364  */
2365
2366 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2367                               struct page *page, u64 start, u64 end,
2368                               int failed_mirror)
2369 {
2370         struct io_failure_record *failrec;
2371         struct inode *inode = page->mapping->host;
2372         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2373         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2374         struct bio *bio;
2375         int read_mode = 0;
2376         blk_status_t status;
2377         int ret;
2378
2379         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2380
2381         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2382         if (ret)
2383                 return ret;
2384
2385         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2386         if (!ret) {
2387                 free_io_failure(failure_tree, tree, failrec);
2388                 return -EIO;
2389         }
2390
2391         if (failed_bio->bi_vcnt > 1)
2392                 read_mode |= REQ_FAILFAST_DEV;
2393
2394         phy_offset >>= inode->i_sb->s_blocksize_bits;
2395         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2396                                       start - page_offset(page),
2397                                       (int)phy_offset, failed_bio->bi_end_io,
2398                                       NULL);
2399         if (!bio) {
2400                 free_io_failure(failure_tree, tree, failrec);
2401                 return -EIO;
2402         }
2403         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2404
2405         btrfs_debug(btrfs_sb(inode->i_sb),
2406                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2407                 read_mode, failrec->this_mirror, failrec->in_validation);
2408
2409         status = tree->ops->submit_bio_hook(tree->private_data, bio, failrec->this_mirror,
2410                                          failrec->bio_flags, 0);
2411         if (status) {
2412                 free_io_failure(failure_tree, tree, failrec);
2413                 bio_put(bio);
2414                 ret = blk_status_to_errno(status);
2415         }
2416
2417         return ret;
2418 }
2419
2420 /* lots and lots of room for performance fixes in the end_bio funcs */
2421
2422 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2423 {
2424         int uptodate = (err == 0);
2425         struct extent_io_tree *tree;
2426         int ret = 0;
2427
2428         tree = &BTRFS_I(page->mapping->host)->io_tree;
2429
2430         if (tree->ops && tree->ops->writepage_end_io_hook)
2431                 tree->ops->writepage_end_io_hook(page, start, end, NULL,
2432                                 uptodate);
2433
2434         if (!uptodate) {
2435                 ClearPageUptodate(page);
2436                 SetPageError(page);
2437                 ret = err < 0 ? err : -EIO;
2438                 mapping_set_error(page->mapping, ret);
2439         }
2440 }
2441
2442 /*
2443  * after a writepage IO is done, we need to:
2444  * clear the uptodate bits on error
2445  * clear the writeback bits in the extent tree for this IO
2446  * end_page_writeback if the page has no more pending IO
2447  *
2448  * Scheduling is not allowed, so the extent state tree is expected
2449  * to have one and only one object corresponding to this IO.
2450  */
2451 static void end_bio_extent_writepage(struct bio *bio)
2452 {
2453         int error = blk_status_to_errno(bio->bi_status);
2454         struct bio_vec *bvec;
2455         u64 start;
2456         u64 end;
2457         int i;
2458
2459         bio_for_each_segment_all(bvec, bio, i) {
2460                 struct page *page = bvec->bv_page;
2461                 struct inode *inode = page->mapping->host;
2462                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2463
2464                 /* We always issue full-page reads, but if some block
2465                  * in a page fails to read, blk_update_request() will
2466                  * advance bv_offset and adjust bv_len to compensate.
2467                  * Print a warning for nonzero offsets, and an error
2468                  * if they don't add up to a full page.  */
2469                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2470                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2471                                 btrfs_err(fs_info,
2472                                    "partial page write in btrfs with offset %u and length %u",
2473                                         bvec->bv_offset, bvec->bv_len);
2474                         else
2475                                 btrfs_info(fs_info,
2476                                    "incomplete page write in btrfs with offset %u and length %u",
2477                                         bvec->bv_offset, bvec->bv_len);
2478                 }
2479
2480                 start = page_offset(page);
2481                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2482
2483                 end_extent_writepage(page, error, start, end);
2484                 end_page_writeback(page);
2485         }
2486
2487         bio_put(bio);
2488 }
2489
2490 static void
2491 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2492                               int uptodate)
2493 {
2494         struct extent_state *cached = NULL;
2495         u64 end = start + len - 1;
2496
2497         if (uptodate && tree->track_uptodate)
2498                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2499         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2500 }
2501
2502 /*
2503  * after a readpage IO is done, we need to:
2504  * clear the uptodate bits on error
2505  * set the uptodate bits if things worked
2506  * set the page up to date if all extents in the tree are uptodate
2507  * clear the lock bit in the extent tree
2508  * unlock the page if there are no other extents locked for it
2509  *
2510  * Scheduling is not allowed, so the extent state tree is expected
2511  * to have one and only one object corresponding to this IO.
2512  */
2513 static void end_bio_extent_readpage(struct bio *bio)
2514 {
2515         struct bio_vec *bvec;
2516         int uptodate = !bio->bi_status;
2517         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2518         struct extent_io_tree *tree, *failure_tree;
2519         u64 offset = 0;
2520         u64 start;
2521         u64 end;
2522         u64 len;
2523         u64 extent_start = 0;
2524         u64 extent_len = 0;
2525         int mirror;
2526         int ret;
2527         int i;
2528
2529         bio_for_each_segment_all(bvec, bio, i) {
2530                 struct page *page = bvec->bv_page;
2531                 struct inode *inode = page->mapping->host;
2532                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2533
2534                 btrfs_debug(fs_info,
2535                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2536                         (u64)bio->bi_iter.bi_sector, bio->bi_status,
2537                         io_bio->mirror_num);
2538                 tree = &BTRFS_I(inode)->io_tree;
2539                 failure_tree = &BTRFS_I(inode)->io_failure_tree;
2540
2541                 /* We always issue full-page reads, but if some block
2542                  * in a page fails to read, blk_update_request() will
2543                  * advance bv_offset and adjust bv_len to compensate.
2544                  * Print a warning for nonzero offsets, and an error
2545                  * if they don't add up to a full page.  */
2546                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2547                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2548                                 btrfs_err(fs_info,
2549                                         "partial page read in btrfs with offset %u and length %u",
2550                                         bvec->bv_offset, bvec->bv_len);
2551                         else
2552                                 btrfs_info(fs_info,
2553                                         "incomplete page read in btrfs with offset %u and length %u",
2554                                         bvec->bv_offset, bvec->bv_len);
2555                 }
2556
2557                 start = page_offset(page);
2558                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2559                 len = bvec->bv_len;
2560
2561                 mirror = io_bio->mirror_num;
2562                 if (likely(uptodate && tree->ops)) {
2563                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2564                                                               page, start, end,
2565                                                               mirror);
2566                         if (ret)
2567                                 uptodate = 0;
2568                         else
2569                                 clean_io_failure(BTRFS_I(inode)->root->fs_info,
2570                                                  failure_tree, tree, start,
2571                                                  page,
2572                                                  btrfs_ino(BTRFS_I(inode)), 0);
2573                 }
2574
2575                 if (likely(uptodate))
2576                         goto readpage_ok;
2577
2578                 if (tree->ops) {
2579                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2580                         if (ret == -EAGAIN) {
2581                                 /*
2582                                  * Data inode's readpage_io_failed_hook() always
2583                                  * returns -EAGAIN.
2584                                  *
2585                                  * The generic bio_readpage_error handles errors
2586                                  * the following way: If possible, new read
2587                                  * requests are created and submitted and will
2588                                  * end up in end_bio_extent_readpage as well (if
2589                                  * we're lucky, not in the !uptodate case). In
2590                                  * that case it returns 0 and we just go on with
2591                                  * the next page in our bio. If it can't handle
2592                                  * the error it will return -EIO and we remain
2593                                  * responsible for that page.
2594                                  */
2595                                 ret = bio_readpage_error(bio, offset, page,
2596                                                          start, end, mirror);
2597                                 if (ret == 0) {
2598                                         uptodate = !bio->bi_status;
2599                                         offset += len;
2600                                         continue;
2601                                 }
2602                         }
2603
2604                         /*
2605                          * metadata's readpage_io_failed_hook() always returns
2606                          * -EIO and fixes nothing.  -EIO is also returned if
2607                          * data inode error could not be fixed.
2608                          */
2609                         ASSERT(ret == -EIO);
2610                 }
2611 readpage_ok:
2612                 if (likely(uptodate)) {
2613                         loff_t i_size = i_size_read(inode);
2614                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2615                         unsigned off;
2616
2617                         /* Zero out the end if this page straddles i_size */
2618                         off = i_size & (PAGE_SIZE-1);
2619                         if (page->index == end_index && off)
2620                                 zero_user_segment(page, off, PAGE_SIZE);
2621                         SetPageUptodate(page);
2622                 } else {
2623                         ClearPageUptodate(page);
2624                         SetPageError(page);
2625                 }
2626                 unlock_page(page);
2627                 offset += len;
2628
2629                 if (unlikely(!uptodate)) {
2630                         if (extent_len) {
2631                                 endio_readpage_release_extent(tree,
2632                                                               extent_start,
2633                                                               extent_len, 1);
2634                                 extent_start = 0;
2635                                 extent_len = 0;
2636                         }
2637                         endio_readpage_release_extent(tree, start,
2638                                                       end - start + 1, 0);
2639                 } else if (!extent_len) {
2640                         extent_start = start;
2641                         extent_len = end + 1 - start;
2642                 } else if (extent_start + extent_len == start) {
2643                         extent_len += end + 1 - start;
2644                 } else {
2645                         endio_readpage_release_extent(tree, extent_start,
2646                                                       extent_len, uptodate);
2647                         extent_start = start;
2648                         extent_len = end + 1 - start;
2649                 }
2650         }
2651
2652         if (extent_len)
2653                 endio_readpage_release_extent(tree, extent_start, extent_len,
2654                                               uptodate);
2655         if (io_bio->end_io)
2656                 io_bio->end_io(io_bio, blk_status_to_errno(bio->bi_status));
2657         bio_put(bio);
2658 }
2659
2660 /*
2661  * Initialize the members up to but not including 'bio'. Use after allocating a
2662  * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
2663  * 'bio' because use of __GFP_ZERO is not supported.
2664  */
2665 static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
2666 {
2667         memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
2668 }
2669
2670 /*
2671  * The following helpers allocate a bio. As it's backed by a bioset, it'll
2672  * never fail.  We're returning a bio right now but you can call btrfs_io_bio
2673  * for the appropriate container_of magic
2674  */
2675 struct bio *btrfs_bio_alloc(struct block_device *bdev, u64 first_byte)
2676 {
2677         struct bio *bio;
2678
2679         bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, btrfs_bioset);
2680         bio->bi_bdev = bdev;
2681         bio->bi_iter.bi_sector = first_byte >> 9;
2682         btrfs_io_bio_init(btrfs_io_bio(bio));
2683         return bio;
2684 }
2685
2686 struct bio *btrfs_bio_clone(struct bio *bio)
2687 {
2688         struct btrfs_io_bio *btrfs_bio;
2689         struct bio *new;
2690
2691         /* Bio allocation backed by a bioset does not fail */
2692         new = bio_clone_fast(bio, GFP_NOFS, btrfs_bioset);
2693         btrfs_bio = btrfs_io_bio(new);
2694         btrfs_io_bio_init(btrfs_bio);
2695         btrfs_bio->iter = bio->bi_iter;
2696         return new;
2697 }
2698
2699 struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
2700 {
2701         struct bio *bio;
2702
2703         /* Bio allocation backed by a bioset does not fail */
2704         bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, btrfs_bioset);
2705         btrfs_io_bio_init(btrfs_io_bio(bio));
2706         return bio;
2707 }
2708
2709 struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
2710 {
2711         struct bio *bio;
2712         struct btrfs_io_bio *btrfs_bio;
2713
2714         /* this will never fail when it's backed by a bioset */
2715         bio = bio_clone_fast(orig, GFP_NOFS, btrfs_bioset);
2716         ASSERT(bio);
2717
2718         btrfs_bio = btrfs_io_bio(bio);
2719         btrfs_io_bio_init(btrfs_bio);
2720
2721         bio_trim(bio, offset >> 9, size >> 9);
2722         btrfs_bio->iter = bio->bi_iter;
2723         return bio;
2724 }
2725
2726 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2727                                        unsigned long bio_flags)
2728 {
2729         blk_status_t ret = 0;
2730         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2731         struct page *page = bvec->bv_page;
2732         struct extent_io_tree *tree = bio->bi_private;
2733         u64 start;
2734
2735         start = page_offset(page) + bvec->bv_offset;
2736
2737         bio->bi_private = NULL;
2738         bio_get(bio);
2739
2740         if (tree->ops)
2741                 ret = tree->ops->submit_bio_hook(tree->private_data, bio,
2742                                            mirror_num, bio_flags, start);
2743         else
2744                 btrfsic_submit_bio(bio);
2745
2746         bio_put(bio);
2747         return blk_status_to_errno(ret);
2748 }
2749
2750 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2751                      unsigned long offset, size_t size, struct bio *bio,
2752                      unsigned long bio_flags)
2753 {
2754         int ret = 0;
2755         if (tree->ops)
2756                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2757                                                 bio_flags);
2758         return ret;
2759
2760 }
2761
2762 static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree,
2763                               struct writeback_control *wbc,
2764                               struct page *page, sector_t sector,
2765                               size_t size, unsigned long offset,
2766                               struct block_device *bdev,
2767                               struct bio **bio_ret,
2768                               bio_end_io_t end_io_func,
2769                               int mirror_num,
2770                               unsigned long prev_bio_flags,
2771                               unsigned long bio_flags,
2772                               bool force_bio_submit)
2773 {
2774         int ret = 0;
2775         struct bio *bio;
2776         int contig = 0;
2777         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2778         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2779
2780         if (bio_ret && *bio_ret) {
2781                 bio = *bio_ret;
2782                 if (old_compressed)
2783                         contig = bio->bi_iter.bi_sector == sector;
2784                 else
2785                         contig = bio_end_sector(bio) == sector;
2786
2787                 if (prev_bio_flags != bio_flags || !contig ||
2788                     force_bio_submit ||
2789                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2790                     bio_add_page(bio, page, page_size, offset) < page_size) {
2791                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2792                         if (ret < 0) {
2793                                 *bio_ret = NULL;
2794                                 return ret;
2795                         }
2796                         bio = NULL;
2797                 } else {
2798                         if (wbc)
2799                                 wbc_account_io(wbc, page, page_size);
2800                         return 0;
2801                 }
2802         }
2803
2804         bio = btrfs_bio_alloc(bdev, sector << 9);
2805         bio_add_page(bio, page, page_size, offset);
2806         bio->bi_end_io = end_io_func;
2807         bio->bi_private = tree;
2808         bio->bi_write_hint = page->mapping->host->i_write_hint;
2809         bio_set_op_attrs(bio, op, op_flags);
2810         if (wbc) {
2811                 wbc_init_bio(wbc, bio);
2812                 wbc_account_io(wbc, page, page_size);
2813         }
2814
2815         if (bio_ret)
2816                 *bio_ret = bio;
2817         else
2818                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2819
2820         return ret;
2821 }
2822
2823 static void attach_extent_buffer_page(struct extent_buffer *eb,
2824                                       struct page *page)
2825 {
2826         if (!PagePrivate(page)) {
2827                 SetPagePrivate(page);
2828                 get_page(page);
2829                 set_page_private(page, (unsigned long)eb);
2830         } else {
2831                 WARN_ON(page->private != (unsigned long)eb);
2832         }
2833 }
2834
2835 void set_page_extent_mapped(struct page *page)
2836 {
2837         if (!PagePrivate(page)) {
2838                 SetPagePrivate(page);
2839                 get_page(page);
2840                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2841         }
2842 }
2843
2844 static struct extent_map *
2845 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2846                  u64 start, u64 len, get_extent_t *get_extent,
2847                  struct extent_map **em_cached)
2848 {
2849         struct extent_map *em;
2850
2851         if (em_cached && *em_cached) {
2852                 em = *em_cached;
2853                 if (extent_map_in_tree(em) && start >= em->start &&
2854                     start < extent_map_end(em)) {
2855                         refcount_inc(&em->refs);
2856                         return em;
2857                 }
2858
2859                 free_extent_map(em);
2860                 *em_cached = NULL;
2861         }
2862
2863         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2864         if (em_cached && !IS_ERR_OR_NULL(em)) {
2865                 BUG_ON(*em_cached);
2866                 refcount_inc(&em->refs);
2867                 *em_cached = em;
2868         }
2869         return em;
2870 }
2871 /*
2872  * basic readpage implementation.  Locked extent state structs are inserted
2873  * into the tree that are removed when the IO is done (by the end_io
2874  * handlers)
2875  * XXX JDM: This needs looking at to ensure proper page locking
2876  * return 0 on success, otherwise return error
2877  */
2878 static int __do_readpage(struct extent_io_tree *tree,
2879                          struct page *page,
2880                          get_extent_t *get_extent,
2881                          struct extent_map **em_cached,
2882                          struct bio **bio, int mirror_num,
2883                          unsigned long *bio_flags, int read_flags,
2884                          u64 *prev_em_start)
2885 {
2886         struct inode *inode = page->mapping->host;
2887         u64 start = page_offset(page);
2888         u64 page_end = start + PAGE_SIZE - 1;
2889         u64 end;
2890         u64 cur = start;
2891         u64 extent_offset;
2892         u64 last_byte = i_size_read(inode);
2893         u64 block_start;
2894         u64 cur_end;
2895         sector_t sector;
2896         struct extent_map *em;
2897         struct block_device *bdev;
2898         int ret = 0;
2899         int nr = 0;
2900         size_t pg_offset = 0;
2901         size_t iosize;
2902         size_t disk_io_size;
2903         size_t blocksize = inode->i_sb->s_blocksize;
2904         unsigned long this_bio_flag = 0;
2905
2906         set_page_extent_mapped(page);
2907
2908         end = page_end;
2909         if (!PageUptodate(page)) {
2910                 if (cleancache_get_page(page) == 0) {
2911                         BUG_ON(blocksize != PAGE_SIZE);
2912                         unlock_extent(tree, start, end);
2913                         goto out;
2914                 }
2915         }
2916
2917         if (page->index == last_byte >> PAGE_SHIFT) {
2918                 char *userpage;
2919                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2920
2921                 if (zero_offset) {
2922                         iosize = PAGE_SIZE - zero_offset;
2923                         userpage = kmap_atomic(page);
2924                         memset(userpage + zero_offset, 0, iosize);
2925                         flush_dcache_page(page);
2926                         kunmap_atomic(userpage);
2927                 }
2928         }
2929         while (cur <= end) {
2930                 bool force_bio_submit = false;
2931
2932                 if (cur >= last_byte) {
2933                         char *userpage;
2934                         struct extent_state *cached = NULL;
2935
2936                         iosize = PAGE_SIZE - pg_offset;
2937                         userpage = kmap_atomic(page);
2938                         memset(userpage + pg_offset, 0, iosize);
2939                         flush_dcache_page(page);
2940                         kunmap_atomic(userpage);
2941                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2942                                             &cached, GFP_NOFS);
2943                         unlock_extent_cached(tree, cur,
2944                                              cur + iosize - 1,
2945                                              &cached, GFP_NOFS);
2946                         break;
2947                 }
2948                 em = __get_extent_map(inode, page, pg_offset, cur,
2949                                       end - cur + 1, get_extent, em_cached);
2950                 if (IS_ERR_OR_NULL(em)) {
2951                         SetPageError(page);
2952                         unlock_extent(tree, cur, end);
2953                         break;
2954                 }
2955                 extent_offset = cur - em->start;
2956                 BUG_ON(extent_map_end(em) <= cur);
2957                 BUG_ON(end < cur);
2958
2959                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2960                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2961                         extent_set_compress_type(&this_bio_flag,
2962                                                  em->compress_type);
2963                 }
2964
2965                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2966                 cur_end = min(extent_map_end(em) - 1, end);
2967                 iosize = ALIGN(iosize, blocksize);
2968                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2969                         disk_io_size = em->block_len;
2970                         sector = em->block_start >> 9;
2971                 } else {
2972                         sector = (em->block_start + extent_offset) >> 9;
2973                         disk_io_size = iosize;
2974                 }
2975                 bdev = em->bdev;
2976                 block_start = em->block_start;
2977                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2978                         block_start = EXTENT_MAP_HOLE;
2979
2980                 /*
2981                  * If we have a file range that points to a compressed extent
2982                  * and it's followed by a consecutive file range that points to
2983                  * to the same compressed extent (possibly with a different
2984                  * offset and/or length, so it either points to the whole extent
2985                  * or only part of it), we must make sure we do not submit a
2986                  * single bio to populate the pages for the 2 ranges because
2987                  * this makes the compressed extent read zero out the pages
2988                  * belonging to the 2nd range. Imagine the following scenario:
2989                  *
2990                  *  File layout
2991                  *  [0 - 8K]                     [8K - 24K]
2992                  *    |                               |
2993                  *    |                               |
2994                  * points to extent X,         points to extent X,
2995                  * offset 4K, length of 8K     offset 0, length 16K
2996                  *
2997                  * [extent X, compressed length = 4K uncompressed length = 16K]
2998                  *
2999                  * If the bio to read the compressed extent covers both ranges,
3000                  * it will decompress extent X into the pages belonging to the
3001                  * first range and then it will stop, zeroing out the remaining
3002                  * pages that belong to the other range that points to extent X.
3003                  * So here we make sure we submit 2 bios, one for the first
3004                  * range and another one for the third range. Both will target
3005                  * the same physical extent from disk, but we can't currently
3006                  * make the compressed bio endio callback populate the pages
3007                  * for both ranges because each compressed bio is tightly
3008                  * coupled with a single extent map, and each range can have
3009                  * an extent map with a different offset value relative to the
3010                  * uncompressed data of our extent and different lengths. This
3011                  * is a corner case so we prioritize correctness over
3012                  * non-optimal behavior (submitting 2 bios for the same extent).
3013                  */
3014                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3015                     prev_em_start && *prev_em_start != (u64)-1 &&
3016                     *prev_em_start != em->orig_start)
3017                         force_bio_submit = true;
3018
3019                 if (prev_em_start)
3020                         *prev_em_start = em->orig_start;
3021
3022                 free_extent_map(em);
3023                 em = NULL;
3024
3025                 /* we've found a hole, just zero and go on */
3026                 if (block_start == EXTENT_MAP_HOLE) {
3027                         char *userpage;
3028                         struct extent_state *cached = NULL;
3029
3030                         userpage = kmap_atomic(page);
3031                         memset(userpage + pg_offset, 0, iosize);
3032                         flush_dcache_page(page);
3033                         kunmap_atomic(userpage);
3034
3035                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3036                                             &cached, GFP_NOFS);
3037                         unlock_extent_cached(tree, cur,
3038                                              cur + iosize - 1,
3039                                              &cached, GFP_NOFS);
3040                         cur = cur + iosize;
3041                         pg_offset += iosize;
3042                         continue;
3043                 }
3044                 /* the get_extent function already copied into the page */
3045                 if (test_range_bit(tree, cur, cur_end,
3046                                    EXTENT_UPTODATE, 1, NULL)) {
3047                         check_page_uptodate(tree, page);
3048                         unlock_extent(tree, cur, cur + iosize - 1);
3049                         cur = cur + iosize;
3050                         pg_offset += iosize;
3051                         continue;
3052                 }
3053                 /* we have an inline extent but it didn't get marked up
3054                  * to date.  Error out
3055                  */
3056                 if (block_start == EXTENT_MAP_INLINE) {
3057                         SetPageError(page);
3058                         unlock_extent(tree, cur, cur + iosize - 1);
3059                         cur = cur + iosize;
3060                         pg_offset += iosize;
3061                         continue;
3062                 }
3063
3064                 ret = submit_extent_page(REQ_OP_READ, read_flags, tree, NULL,
3065                                          page, sector, disk_io_size, pg_offset,
3066                                          bdev, bio,
3067                                          end_bio_extent_readpage, mirror_num,
3068                                          *bio_flags,
3069                                          this_bio_flag,
3070                                          force_bio_submit);
3071                 if (!ret) {
3072                         nr++;
3073                         *bio_flags = this_bio_flag;
3074                 } else {
3075                         SetPageError(page);
3076                         unlock_extent(tree, cur, cur + iosize - 1);
3077                         goto out;
3078                 }
3079                 cur = cur + iosize;
3080                 pg_offset += iosize;
3081         }
3082 out:
3083         if (!nr) {
3084                 if (!PageError(page))
3085                         SetPageUptodate(page);
3086                 unlock_page(page);
3087         }
3088         return ret;
3089 }
3090
3091 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3092                                              struct page *pages[], int nr_pages,
3093                                              u64 start, u64 end,
3094                                              get_extent_t *get_extent,
3095                                              struct extent_map **em_cached,
3096                                              struct bio **bio, int mirror_num,
3097                                              unsigned long *bio_flags,
3098                                              u64 *prev_em_start)
3099 {
3100         struct inode *inode;
3101         struct btrfs_ordered_extent *ordered;
3102         int index;
3103
3104         inode = pages[0]->mapping->host;
3105         while (1) {
3106                 lock_extent(tree, start, end);
3107                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3108                                                      end - start + 1);
3109                 if (!ordered)
3110                         break;
3111                 unlock_extent(tree, start, end);
3112                 btrfs_start_ordered_extent(inode, ordered, 1);
3113                 btrfs_put_ordered_extent(ordered);
3114         }
3115
3116         for (index = 0; index < nr_pages; index++) {
3117                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3118                               mirror_num, bio_flags, 0, prev_em_start);
3119                 put_page(pages[index]);
3120         }
3121 }
3122
3123 static void __extent_readpages(struct extent_io_tree *tree,
3124                                struct page *pages[],
3125                                int nr_pages, get_extent_t *get_extent,
3126                                struct extent_map **em_cached,
3127                                struct bio **bio, int mirror_num,
3128                                unsigned long *bio_flags,
3129                                u64 *prev_em_start)
3130 {
3131         u64 start = 0;
3132         u64 end = 0;
3133         u64 page_start;
3134         int index;
3135         int first_index = 0;
3136
3137         for (index = 0; index < nr_pages; index++) {
3138                 page_start = page_offset(pages[index]);
3139                 if (!end) {
3140                         start = page_start;
3141                         end = start + PAGE_SIZE - 1;
3142                         first_index = index;
3143                 } else if (end + 1 == page_start) {
3144                         end += PAGE_SIZE;
3145                 } else {
3146                         __do_contiguous_readpages(tree, &pages[first_index],
3147                                                   index - first_index, start,
3148                                                   end, get_extent, em_cached,
3149                                                   bio, mirror_num, bio_flags,
3150                                                   prev_em_start);
3151                         start = page_start;
3152                         end = start + PAGE_SIZE - 1;
3153                         first_index = index;
3154                 }
3155         }
3156
3157         if (end)
3158                 __do_contiguous_readpages(tree, &pages[first_index],
3159                                           index - first_index, start,
3160                                           end, get_extent, em_cached, bio,
3161                                           mirror_num, bio_flags,
3162                                           prev_em_start);
3163 }
3164
3165 static int __extent_read_full_page(struct extent_io_tree *tree,
3166                                    struct page *page,
3167                                    get_extent_t *get_extent,
3168                                    struct bio **bio, int mirror_num,
3169                                    unsigned long *bio_flags, int read_flags)
3170 {
3171         struct inode *inode = page->mapping->host;
3172         struct btrfs_ordered_extent *ordered;
3173         u64 start = page_offset(page);
3174         u64 end = start + PAGE_SIZE - 1;
3175         int ret;
3176
3177         while (1) {
3178                 lock_extent(tree, start, end);
3179                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3180                                                 PAGE_SIZE);
3181                 if (!ordered)
3182                         break;
3183                 unlock_extent(tree, start, end);
3184                 btrfs_start_ordered_extent(inode, ordered, 1);
3185                 btrfs_put_ordered_extent(ordered);
3186         }
3187
3188         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3189                             bio_flags, read_flags, NULL);
3190         return ret;
3191 }
3192
3193 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3194                             get_extent_t *get_extent, int mirror_num)
3195 {
3196         struct bio *bio = NULL;
3197         unsigned long bio_flags = 0;
3198         int ret;
3199
3200         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3201                                       &bio_flags, 0);
3202         if (bio)
3203                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3204         return ret;
3205 }
3206
3207 static void update_nr_written(struct writeback_control *wbc,
3208                               unsigned long nr_written)
3209 {
3210         wbc->nr_to_write -= nr_written;
3211 }
3212
3213 /*
3214  * helper for __extent_writepage, doing all of the delayed allocation setup.
3215  *
3216  * This returns 1 if our fill_delalloc function did all the work required
3217  * to write the page (copy into inline extent).  In this case the IO has
3218  * been started and the page is already unlocked.
3219  *
3220  * This returns 0 if all went well (page still locked)
3221  * This returns < 0 if there were errors (page still locked)
3222  */
3223 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3224                               struct page *page, struct writeback_control *wbc,
3225                               struct extent_page_data *epd,
3226                               u64 delalloc_start,
3227                               unsigned long *nr_written)
3228 {
3229         struct extent_io_tree *tree = epd->tree;
3230         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3231         u64 nr_delalloc;
3232         u64 delalloc_to_write = 0;
3233         u64 delalloc_end = 0;
3234         int ret;
3235         int page_started = 0;
3236
3237         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3238                 return 0;
3239
3240         while (delalloc_end < page_end) {
3241                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3242                                                page,
3243                                                &delalloc_start,
3244                                                &delalloc_end,
3245                                                BTRFS_MAX_EXTENT_SIZE);
3246                 if (nr_delalloc == 0) {
3247                         delalloc_start = delalloc_end + 1;
3248                         continue;
3249                 }
3250                 ret = tree->ops->fill_delalloc(inode, page,
3251                                                delalloc_start,
3252                                                delalloc_end,
3253                                                &page_started,
3254                                                nr_written);
3255                 /* File system has been set read-only */
3256                 if (ret) {
3257                         SetPageError(page);
3258                         /* fill_delalloc should be return < 0 for error
3259                          * but just in case, we use > 0 here meaning the
3260                          * IO is started, so we don't want to return > 0
3261                          * unless things are going well.
3262                          */
3263                         ret = ret < 0 ? ret : -EIO;
3264                         goto done;
3265                 }
3266                 /*
3267                  * delalloc_end is already one less than the total length, so
3268                  * we don't subtract one from PAGE_SIZE
3269                  */
3270                 delalloc_to_write += (delalloc_end - delalloc_start +
3271                                       PAGE_SIZE) >> PAGE_SHIFT;
3272                 delalloc_start = delalloc_end + 1;
3273         }
3274         if (wbc->nr_to_write < delalloc_to_write) {
3275                 int thresh = 8192;
3276
3277                 if (delalloc_to_write < thresh * 2)
3278                         thresh = delalloc_to_write;
3279                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3280                                          thresh);
3281         }
3282
3283         /* did the fill delalloc function already unlock and start
3284          * the IO?
3285          */
3286         if (page_started) {
3287                 /*
3288                  * we've unlocked the page, so we can't update
3289                  * the mapping's writeback index, just update
3290                  * nr_to_write.
3291                  */
3292                 wbc->nr_to_write -= *nr_written;
3293                 return 1;
3294         }
3295
3296         ret = 0;
3297
3298 done:
3299         return ret;
3300 }
3301
3302 /*
3303  * helper for __extent_writepage.  This calls the writepage start hooks,
3304  * and does the loop to map the page into extents and bios.
3305  *
3306  * We return 1 if the IO is started and the page is unlocked,
3307  * 0 if all went well (page still locked)
3308  * < 0 if there were errors (page still locked)
3309  */
3310 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3311                                  struct page *page,
3312                                  struct writeback_control *wbc,
3313                                  struct extent_page_data *epd,
3314                                  loff_t i_size,
3315                                  unsigned long nr_written,
3316                                  int write_flags, int *nr_ret)
3317 {
3318         struct extent_io_tree *tree = epd->tree;
3319         u64 start = page_offset(page);
3320         u64 page_end = start + PAGE_SIZE - 1;
3321         u64 end;
3322         u64 cur = start;
3323         u64 extent_offset;
3324         u64 block_start;
3325         u64 iosize;
3326         sector_t sector;
3327         struct extent_map *em;
3328         struct block_device *bdev;
3329         size_t pg_offset = 0;
3330         size_t blocksize;
3331         int ret = 0;
3332         int nr = 0;
3333         bool compressed;
3334
3335         if (tree->ops && tree->ops->writepage_start_hook) {
3336                 ret = tree->ops->writepage_start_hook(page, start,
3337                                                       page_end);
3338                 if (ret) {
3339                         /* Fixup worker will requeue */
3340                         if (ret == -EBUSY)
3341                                 wbc->pages_skipped++;
3342                         else
3343                                 redirty_page_for_writepage(wbc, page);
3344
3345                         update_nr_written(wbc, nr_written);
3346                         unlock_page(page);
3347                         return 1;
3348                 }
3349         }
3350
3351         /*
3352          * we don't want to touch the inode after unlocking the page,
3353          * so we update the mapping writeback index now
3354          */
3355         update_nr_written(wbc, nr_written + 1);
3356
3357         end = page_end;
3358         if (i_size <= start) {
3359                 if (tree->ops && tree->ops->writepage_end_io_hook)
3360                         tree->ops->writepage_end_io_hook(page, start,
3361                                                          page_end, NULL, 1);
3362                 goto done;
3363         }
3364
3365         blocksize = inode->i_sb->s_blocksize;
3366
3367         while (cur <= end) {
3368                 u64 em_end;
3369
3370                 if (cur >= i_size) {
3371                         if (tree->ops && tree->ops->writepage_end_io_hook)
3372                                 tree->ops->writepage_end_io_hook(page, cur,
3373                                                          page_end, NULL, 1);
3374                         break;
3375                 }
3376                 em = epd->get_extent(BTRFS_I(inode), page, pg_offset, cur,
3377                                      end - cur + 1, 1);
3378                 if (IS_ERR_OR_NULL(em)) {
3379                         SetPageError(page);
3380                         ret = PTR_ERR_OR_ZERO(em);
3381                         break;
3382                 }
3383
3384                 extent_offset = cur - em->start;
3385                 em_end = extent_map_end(em);
3386                 BUG_ON(em_end <= cur);
3387                 BUG_ON(end < cur);
3388                 iosize = min(em_end - cur, end - cur + 1);
3389                 iosize = ALIGN(iosize, blocksize);
3390                 sector = (em->block_start + extent_offset) >> 9;
3391                 bdev = em->bdev;
3392                 block_start = em->block_start;
3393                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3394                 free_extent_map(em);
3395                 em = NULL;
3396
3397                 /*
3398                  * compressed and inline extents are written through other
3399                  * paths in the FS
3400                  */
3401                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3402                     block_start == EXTENT_MAP_INLINE) {
3403                         /*
3404                          * end_io notification does not happen here for
3405                          * compressed extents
3406                          */
3407                         if (!compressed && tree->ops &&
3408                             tree->ops->writepage_end_io_hook)
3409                                 tree->ops->writepage_end_io_hook(page, cur,
3410                                                          cur + iosize - 1,
3411                                                          NULL, 1);
3412                         else if (compressed) {
3413                                 /* we don't want to end_page_writeback on
3414                                  * a compressed extent.  this happens
3415                                  * elsewhere
3416                                  */
3417                                 nr++;
3418                         }
3419
3420                         cur += iosize;
3421                         pg_offset += iosize;
3422                         continue;
3423                 }
3424
3425                 set_range_writeback(tree, cur, cur + iosize - 1);
3426                 if (!PageWriteback(page)) {
3427                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3428                                    "page %lu not writeback, cur %llu end %llu",
3429                                page->index, cur, end);
3430                 }
3431
3432                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3433                                          page, sector, iosize, pg_offset,
3434                                          bdev, &epd->bio,
3435                                          end_bio_extent_writepage,
3436                                          0, 0, 0, false);
3437                 if (ret) {
3438                         SetPageError(page);
3439                         if (PageWriteback(page))
3440                                 end_page_writeback(page);
3441                 }
3442
3443                 cur = cur + iosize;
3444                 pg_offset += iosize;
3445                 nr++;
3446         }
3447 done:
3448         *nr_ret = nr;
3449         return ret;
3450 }
3451
3452 /*
3453  * the writepage semantics are similar to regular writepage.  extent
3454  * records are inserted to lock ranges in the tree, and as dirty areas
3455  * are found, they are marked writeback.  Then the lock bits are removed
3456  * and the end_io handler clears the writeback ranges
3457  */
3458 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3459                               void *data)
3460 {
3461         struct inode *inode = page->mapping->host;
3462         struct extent_page_data *epd = data;
3463         u64 start = page_offset(page);
3464         u64 page_end = start + PAGE_SIZE - 1;
3465         int ret;
3466         int nr = 0;
3467         size_t pg_offset = 0;
3468         loff_t i_size = i_size_read(inode);
3469         unsigned long end_index = i_size >> PAGE_SHIFT;
3470         int write_flags = 0;
3471         unsigned long nr_written = 0;
3472
3473         if (wbc->sync_mode == WB_SYNC_ALL)
3474                 write_flags = REQ_SYNC;
3475
3476         trace___extent_writepage(page, inode, wbc);
3477
3478         WARN_ON(!PageLocked(page));
3479
3480         ClearPageError(page);
3481
3482         pg_offset = i_size & (PAGE_SIZE - 1);
3483         if (page->index > end_index ||
3484            (page->index == end_index && !pg_offset)) {
3485                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3486                 unlock_page(page);
3487                 return 0;
3488         }
3489
3490         if (page->index == end_index) {
3491                 char *userpage;
3492
3493                 userpage = kmap_atomic(page);
3494                 memset(userpage + pg_offset, 0,
3495                        PAGE_SIZE - pg_offset);
3496                 kunmap_atomic(userpage);
3497                 flush_dcache_page(page);
3498         }
3499
3500         pg_offset = 0;
3501
3502         set_page_extent_mapped(page);
3503
3504         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3505         if (ret == 1)
3506                 goto done_unlocked;
3507         if (ret)
3508                 goto done;
3509
3510         ret = __extent_writepage_io(inode, page, wbc, epd,
3511                                     i_size, nr_written, write_flags, &nr);
3512         if (ret == 1)
3513                 goto done_unlocked;
3514
3515 done:
3516         if (nr == 0) {
3517                 /* make sure the mapping tag for page dirty gets cleared */
3518                 set_page_writeback(page);
3519                 end_page_writeback(page);
3520         }
3521         if (PageError(page)) {
3522                 ret = ret < 0 ? ret : -EIO;
3523                 end_extent_writepage(page, ret, start, page_end);
3524         }
3525         unlock_page(page);
3526         return ret;
3527
3528 done_unlocked:
3529         return 0;
3530 }
3531
3532 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3533 {
3534         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3535                        TASK_UNINTERRUPTIBLE);
3536 }
3537
3538 static noinline_for_stack int
3539 lock_extent_buffer_for_io(struct extent_buffer *eb,
3540                           struct btrfs_fs_info *fs_info,
3541                           struct extent_page_data *epd)
3542 {
3543         unsigned long i, num_pages;
3544         int flush = 0;
3545         int ret = 0;
3546
3547         if (!btrfs_try_tree_write_lock(eb)) {
3548                 flush = 1;
3549                 flush_write_bio(epd);
3550                 btrfs_tree_lock(eb);
3551         }
3552
3553         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3554                 btrfs_tree_unlock(eb);
3555                 if (!epd->sync_io)
3556                         return 0;
3557                 if (!flush) {
3558                         flush_write_bio(epd);
3559                         flush = 1;
3560                 }
3561                 while (1) {
3562                         wait_on_extent_buffer_writeback(eb);
3563                         btrfs_tree_lock(eb);
3564                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3565                                 break;
3566                         btrfs_tree_unlock(eb);
3567                 }
3568         }
3569
3570         /*
3571          * We need to do this to prevent races in people who check if the eb is
3572          * under IO since we can end up having no IO bits set for a short period
3573          * of time.
3574          */
3575         spin_lock(&eb->refs_lock);
3576         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3577                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3578                 spin_unlock(&eb->refs_lock);
3579                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3580                 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
3581                                          -eb->len,
3582                                          fs_info->dirty_metadata_batch);
3583                 ret = 1;
3584         } else {
3585                 spin_unlock(&eb->refs_lock);
3586         }
3587
3588         btrfs_tree_unlock(eb);
3589
3590         if (!ret)
3591                 return ret;
3592
3593         num_pages = num_extent_pages(eb->start, eb->len);
3594         for (i = 0; i < num_pages; i++) {
3595                 struct page *p = eb->pages[i];
3596
3597                 if (!trylock_page(p)) {
3598                         if (!flush) {
3599                                 flush_write_bio(epd);
3600                                 flush = 1;
3601                         }
3602                         lock_page(p);
3603                 }
3604         }
3605
3606         return ret;
3607 }
3608
3609 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3610 {
3611         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3612         smp_mb__after_atomic();
3613         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3614 }
3615
3616 static void set_btree_ioerr(struct page *page)
3617 {
3618         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3619
3620         SetPageError(page);
3621         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3622                 return;
3623
3624         /*
3625          * If writeback for a btree extent that doesn't belong to a log tree
3626          * failed, increment the counter transaction->eb_write_errors.
3627          * We do this because while the transaction is running and before it's
3628          * committing (when we call filemap_fdata[write|wait]_range against
3629          * the btree inode), we might have
3630          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3631          * returns an error or an error happens during writeback, when we're
3632          * committing the transaction we wouldn't know about it, since the pages
3633          * can be no longer dirty nor marked anymore for writeback (if a
3634          * subsequent modification to the extent buffer didn't happen before the
3635          * transaction commit), which makes filemap_fdata[write|wait]_range not
3636          * able to find the pages tagged with SetPageError at transaction
3637          * commit time. So if this happens we must abort the transaction,
3638          * otherwise we commit a super block with btree roots that point to
3639          * btree nodes/leafs whose content on disk is invalid - either garbage
3640          * or the content of some node/leaf from a past generation that got
3641          * cowed or deleted and is no longer valid.
3642          *
3643          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3644          * not be enough - we need to distinguish between log tree extents vs
3645          * non-log tree extents, and the next filemap_fdatawait_range() call
3646          * will catch and clear such errors in the mapping - and that call might
3647          * be from a log sync and not from a transaction commit. Also, checking
3648          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3649          * not done and would not be reliable - the eb might have been released
3650          * from memory and reading it back again means that flag would not be
3651          * set (since it's a runtime flag, not persisted on disk).
3652          *
3653          * Using the flags below in the btree inode also makes us achieve the
3654          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3655          * writeback for all dirty pages and before filemap_fdatawait_range()
3656          * is called, the writeback for all dirty pages had already finished
3657          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3658          * filemap_fdatawait_range() would return success, as it could not know
3659          * that writeback errors happened (the pages were no longer tagged for
3660          * writeback).
3661          */
3662         switch (eb->log_index) {
3663         case -1:
3664                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3665                 break;
3666         case 0:
3667                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3668                 break;
3669         case 1:
3670                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3671                 break;
3672         default:
3673                 BUG(); /* unexpected, logic error */
3674         }
3675 }
3676
3677 static void end_bio_extent_buffer_writepage(struct bio *bio)
3678 {
3679         struct bio_vec *bvec;
3680         struct extent_buffer *eb;
3681         int i, done;
3682
3683         bio_for_each_segment_all(bvec, bio, i) {
3684                 struct page *page = bvec->bv_page;
3685
3686                 eb = (struct extent_buffer *)page->private;
3687                 BUG_ON(!eb);
3688                 done = atomic_dec_and_test(&eb->io_pages);
3689
3690                 if (bio->bi_status ||
3691                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3692                         ClearPageUptodate(page);
3693                         set_btree_ioerr(page);
3694                 }
3695
3696                 end_page_writeback(page);
3697
3698                 if (!done)
3699                         continue;
3700
3701                 end_extent_buffer_writeback(eb);
3702         }
3703
3704         bio_put(bio);
3705 }
3706
3707 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3708                         struct btrfs_fs_info *fs_info,
3709                         struct writeback_control *wbc,
3710                         struct extent_page_data *epd)
3711 {
3712         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3713         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3714         u64 offset = eb->start;
3715         u32 nritems;
3716         unsigned long i, num_pages;
3717         unsigned long bio_flags = 0;
3718         unsigned long start, end;
3719         int write_flags = (epd->sync_io ? REQ_SYNC : 0) | REQ_META;
3720         int ret = 0;
3721
3722         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3723         num_pages = num_extent_pages(eb->start, eb->len);
3724         atomic_set(&eb->io_pages, num_pages);
3725         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3726                 bio_flags = EXTENT_BIO_TREE_LOG;
3727
3728         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3729         nritems = btrfs_header_nritems(eb);
3730         if (btrfs_header_level(eb) > 0) {
3731                 end = btrfs_node_key_ptr_offset(nritems);
3732
3733                 memzero_extent_buffer(eb, end, eb->len - end);
3734         } else {
3735                 /*
3736                  * leaf:
3737                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3738                  */
3739                 start = btrfs_item_nr_offset(nritems);
3740                 end = BTRFS_LEAF_DATA_OFFSET + leaf_data_end(fs_info, eb);
3741                 memzero_extent_buffer(eb, start, end - start);
3742         }
3743
3744         for (i = 0; i < num_pages; i++) {
3745                 struct page *p = eb->pages[i];
3746
3747                 clear_page_dirty_for_io(p);
3748                 set_page_writeback(p);
3749                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3750                                          p, offset >> 9, PAGE_SIZE, 0, bdev,
3751                                          &epd->bio,
3752                                          end_bio_extent_buffer_writepage,
3753                                          0, epd->bio_flags, bio_flags, false);
3754                 epd->bio_flags = bio_flags;
3755                 if (ret) {
3756                         set_btree_ioerr(p);
3757                         if (PageWriteback(p))
3758                                 end_page_writeback(p);
3759                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3760                                 end_extent_buffer_writeback(eb);
3761                         ret = -EIO;
3762                         break;
3763                 }
3764                 offset += PAGE_SIZE;
3765                 update_nr_written(wbc, 1);
3766                 unlock_page(p);
3767         }
3768
3769         if (unlikely(ret)) {
3770                 for (; i < num_pages; i++) {
3771                         struct page *p = eb->pages[i];
3772                         clear_page_dirty_for_io(p);
3773                         unlock_page(p);
3774                 }
3775         }
3776
3777         return ret;
3778 }
3779
3780 int btree_write_cache_pages(struct address_space *mapping,
3781                                    struct writeback_control *wbc)
3782 {
3783         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3784         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3785         struct extent_buffer *eb, *prev_eb = NULL;
3786         struct extent_page_data epd = {
3787                 .bio = NULL,
3788                 .tree = tree,
3789                 .extent_locked = 0,
3790                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3791                 .bio_flags = 0,
3792         };
3793         int ret = 0;
3794         int done = 0;
3795         int nr_to_write_done = 0;
3796         struct pagevec pvec;
3797         int nr_pages;
3798         pgoff_t index;
3799         pgoff_t end;            /* Inclusive */
3800         int scanned = 0;
3801         int tag;
3802
3803         pagevec_init(&pvec, 0);
3804         if (wbc->range_cyclic) {
3805                 index = mapping->writeback_index; /* Start from prev offset */
3806                 end = -1;
3807         } else {
3808                 index = wbc->range_start >> PAGE_SHIFT;
3809                 end = wbc->range_end >> PAGE_SHIFT;
3810                 scanned = 1;
3811         }
3812         if (wbc->sync_mode == WB_SYNC_ALL)
3813                 tag = PAGECACHE_TAG_TOWRITE;
3814         else
3815                 tag = PAGECACHE_TAG_DIRTY;
3816 retry:
3817         if (wbc->sync_mode == WB_SYNC_ALL)
3818                 tag_pages_for_writeback(mapping, index, end);
3819         while (!done && !nr_to_write_done && (index <= end) &&
3820                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3821                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3822                 unsigned i;
3823
3824                 scanned = 1;
3825                 for (i = 0; i < nr_pages; i++) {
3826                         struct page *page = pvec.pages[i];
3827
3828                         if (!PagePrivate(page))
3829                                 continue;
3830
3831                         if (!wbc->range_cyclic && page->index > end) {
3832                                 done = 1;
3833                                 break;
3834                         }
3835
3836                         spin_lock(&mapping->private_lock);
3837                         if (!PagePrivate(page)) {
3838                                 spin_unlock(&mapping->private_lock);
3839                                 continue;
3840                         }
3841
3842                         eb = (struct extent_buffer *)page->private;
3843
3844                         /*
3845                          * Shouldn't happen and normally this would be a BUG_ON
3846                          * but no sense in crashing the users box for something
3847                          * we can survive anyway.
3848                          */
3849                         if (WARN_ON(!eb)) {
3850                                 spin_unlock(&mapping->private_lock);
3851                                 continue;
3852                         }
3853
3854                         if (eb == prev_eb) {
3855                                 spin_unlock(&mapping->private_lock);
3856                                 continue;
3857                         }
3858
3859                         ret = atomic_inc_not_zero(&eb->refs);
3860                         spin_unlock(&mapping->private_lock);
3861                         if (!ret)
3862                                 continue;
3863
3864                         prev_eb = eb;
3865                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3866                         if (!ret) {
3867                                 free_extent_buffer(eb);
3868                                 continue;
3869                         }
3870
3871                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3872                         if (ret) {
3873                                 done = 1;
3874                                 free_extent_buffer(eb);
3875                                 break;
3876                         }
3877                         free_extent_buffer(eb);
3878
3879                         /*
3880                          * the filesystem may choose to bump up nr_to_write.
3881                          * We have to make sure to honor the new nr_to_write
3882                          * at any time
3883                          */
3884                         nr_to_write_done = wbc->nr_to_write <= 0;
3885                 }
3886                 pagevec_release(&pvec);
3887                 cond_resched();
3888         }
3889         if (!scanned && !done) {
3890                 /*
3891                  * We hit the last page and there is more work to be done: wrap
3892                  * back to the start of the file
3893                  */
3894                 scanned = 1;
3895                 index = 0;
3896                 goto retry;
3897         }
3898         flush_write_bio(&epd);
3899         return ret;
3900 }
3901
3902 /**
3903  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3904  * @mapping: address space structure to write
3905  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3906  * @writepage: function called for each page
3907  * @data: data passed to writepage function
3908  *
3909  * If a page is already under I/O, write_cache_pages() skips it, even
3910  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3911  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3912  * and msync() need to guarantee that all the data which was dirty at the time
3913  * the call was made get new I/O started against them.  If wbc->sync_mode is
3914  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3915  * existing IO to complete.
3916  */
3917 static int extent_write_cache_pages(struct address_space *mapping,
3918                              struct writeback_control *wbc,
3919                              writepage_t writepage, void *data,
3920                              void (*flush_fn)(void *))
3921 {
3922         struct inode *inode = mapping->host;
3923         int ret = 0;
3924         int done = 0;
3925         int nr_to_write_done = 0;
3926         struct pagevec pvec;
3927         int nr_pages;
3928         pgoff_t index;
3929         pgoff_t end;            /* Inclusive */
3930         pgoff_t done_index;
3931         int range_whole = 0;
3932         int scanned = 0;
3933         int tag;
3934
3935         /*
3936          * We have to hold onto the inode so that ordered extents can do their
3937          * work when the IO finishes.  The alternative to this is failing to add
3938          * an ordered extent if the igrab() fails there and that is a huge pain
3939          * to deal with, so instead just hold onto the inode throughout the
3940          * writepages operation.  If it fails here we are freeing up the inode
3941          * anyway and we'd rather not waste our time writing out stuff that is
3942          * going to be truncated anyway.
3943          */
3944         if (!igrab(inode))
3945                 return 0;
3946
3947         pagevec_init(&pvec, 0);
3948         if (wbc->range_cyclic) {
3949                 index = mapping->writeback_index; /* Start from prev offset */
3950                 end = -1;
3951         } else {
3952                 index = wbc->range_start >> PAGE_SHIFT;
3953                 end = wbc->range_end >> PAGE_SHIFT;
3954                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3955                         range_whole = 1;
3956                 scanned = 1;
3957         }
3958         if (wbc->sync_mode == WB_SYNC_ALL)
3959                 tag = PAGECACHE_TAG_TOWRITE;
3960         else
3961                 tag = PAGECACHE_TAG_DIRTY;
3962 retry:
3963         if (wbc->sync_mode == WB_SYNC_ALL)
3964                 tag_pages_for_writeback(mapping, index, end);
3965         done_index = index;
3966         while (!done && !nr_to_write_done && (index <= end) &&
3967                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3968                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3969                 unsigned i;
3970
3971                 scanned = 1;
3972                 for (i = 0; i < nr_pages; i++) {
3973                         struct page *page = pvec.pages[i];
3974
3975                         done_index = page->index;
3976                         /*
3977                          * At this point we hold neither mapping->tree_lock nor
3978                          * lock on the page itself: the page may be truncated or
3979                          * invalidated (changing page->mapping to NULL), or even
3980                          * swizzled back from swapper_space to tmpfs file
3981                          * mapping
3982                          */
3983                         if (!trylock_page(page)) {
3984                                 flush_fn(data);
3985                                 lock_page(page);
3986                         }
3987
3988                         if (unlikely(page->mapping != mapping)) {
3989                                 unlock_page(page);
3990                                 continue;
3991                         }
3992
3993                         if (!wbc->range_cyclic && page->index > end) {
3994                                 done = 1;
3995                                 unlock_page(page);
3996                                 continue;
3997                         }
3998
3999                         if (wbc->sync_mode != WB_SYNC_NONE) {
4000                                 if (PageWriteback(page))
4001                                         flush_fn(data);
4002                                 wait_on_page_writeback(page);
4003                         }
4004
4005                         if (PageWriteback(page) ||
4006                             !clear_page_dirty_for_io(page)) {
4007                                 unlock_page(page);
4008                                 continue;
4009                         }
4010
4011                         ret = (*writepage)(page, wbc, data);
4012
4013                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4014                                 unlock_page(page);
4015                                 ret = 0;
4016                         }
4017                         if (ret < 0) {
4018                                 /*
4019                                  * done_index is set past this page,
4020                                  * so media errors will not choke
4021                                  * background writeout for the entire
4022                                  * file. This has consequences for
4023                                  * range_cyclic semantics (ie. it may
4024                                  * not be suitable for data integrity
4025                                  * writeout).
4026                                  */
4027                                 done_index = page->index + 1;
4028                                 done = 1;
4029                                 break;
4030                         }
4031
4032                         /*
4033                          * the filesystem may choose to bump up nr_to_write.
4034                          * We have to make sure to honor the new nr_to_write
4035                          * at any time
4036                          */
4037                         nr_to_write_done = wbc->nr_to_write <= 0;
4038                 }
4039                 pagevec_release(&pvec);
4040                 cond_resched();
4041         }
4042         if (!scanned && !done) {
4043                 /*
4044                  * We hit the last page and there is more work to be done: wrap
4045                  * back to the start of the file
4046                  */
4047                 scanned = 1;
4048                 index = 0;
4049                 goto retry;
4050         }
4051
4052         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4053                 mapping->writeback_index = done_index;
4054
4055         btrfs_add_delayed_iput(inode);
4056         return ret;
4057 }
4058
4059 static void flush_epd_write_bio(struct extent_page_data *epd)
4060 {
4061         if (epd->bio) {
4062                 int ret;
4063
4064                 bio_set_op_attrs(epd->bio, REQ_OP_WRITE,
4065                                  epd->sync_io ? REQ_SYNC : 0);
4066
4067                 ret = submit_one_bio(epd->bio, 0, epd->bio_flags);
4068                 BUG_ON(ret < 0); /* -ENOMEM */
4069                 epd->bio = NULL;
4070         }
4071 }
4072
4073 static noinline void flush_write_bio(void *data)
4074 {
4075         struct extent_page_data *epd = data;
4076         flush_epd_write_bio(epd);
4077 }
4078
4079 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4080                           get_extent_t *get_extent,
4081                           struct writeback_control *wbc)
4082 {
4083         int ret;
4084         struct extent_page_data epd = {
4085                 .bio = NULL,
4086                 .tree = tree,
4087                 .get_extent = get_extent,
4088                 .extent_locked = 0,
4089                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4090                 .bio_flags = 0,
4091         };
4092
4093         ret = __extent_writepage(page, wbc, &epd);
4094
4095         flush_epd_write_bio(&epd);
4096         return ret;
4097 }
4098
4099 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4100                               u64 start, u64 end, get_extent_t *get_extent,
4101                               int mode)
4102 {
4103         int ret = 0;
4104         struct address_space *mapping = inode->i_mapping;
4105         struct page *page;
4106         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4107                 PAGE_SHIFT;
4108
4109         struct extent_page_data epd = {
4110                 .bio = NULL,
4111                 .tree = tree,
4112                 .get_extent = get_extent,
4113                 .extent_locked = 1,
4114                 .sync_io = mode == WB_SYNC_ALL,
4115                 .bio_flags = 0,
4116         };
4117         struct writeback_control wbc_writepages = {
4118                 .sync_mode      = mode,
4119                 .nr_to_write    = nr_pages * 2,
4120                 .range_start    = start,
4121                 .range_end      = end + 1,
4122         };
4123
4124         while (start <= end) {
4125                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4126                 if (clear_page_dirty_for_io(page))
4127                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4128                 else {
4129                         if (tree->ops && tree->ops->writepage_end_io_hook)
4130                                 tree->ops->writepage_end_io_hook(page, start,
4131                                                  start + PAGE_SIZE - 1,
4132                                                  NULL, 1);
4133                         unlock_page(page);
4134                 }
4135                 put_page(page);
4136                 start += PAGE_SIZE;
4137         }
4138
4139         flush_epd_write_bio(&epd);
4140         return ret;
4141 }
4142
4143 int extent_writepages(struct extent_io_tree *tree,
4144                       struct address_space *mapping,
4145                       get_extent_t *get_extent,
4146                       struct writeback_control *wbc)
4147 {
4148         int ret = 0;
4149         struct extent_page_data epd = {
4150                 .bio = NULL,
4151                 .tree = tree,
4152                 .get_extent = get_extent,
4153                 .extent_locked = 0,
4154                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4155                 .bio_flags = 0,
4156         };
4157
4158         ret = extent_write_cache_pages(mapping, wbc, __extent_writepage, &epd,
4159                                        flush_write_bio);
4160         flush_epd_write_bio(&epd);
4161         return ret;
4162 }
4163
4164 int extent_readpages(struct extent_io_tree *tree,
4165                      struct address_space *mapping,
4166                      struct list_head *pages, unsigned nr_pages,
4167                      get_extent_t get_extent)
4168 {
4169         struct bio *bio = NULL;
4170         unsigned page_idx;
4171         unsigned long bio_flags = 0;
4172         struct page *pagepool[16];
4173         struct page *page;
4174         struct extent_map *em_cached = NULL;
4175         int nr = 0;
4176         u64 prev_em_start = (u64)-1;
4177
4178         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4179                 page = list_entry(pages->prev, struct page, lru);
4180
4181                 prefetchw(&page->flags);
4182                 list_del(&page->lru);
4183                 if (add_to_page_cache_lru(page, mapping,
4184                                         page->index,
4185                                         readahead_gfp_mask(mapping))) {
4186                         put_page(page);
4187                         continue;
4188                 }
4189
4190                 pagepool[nr++] = page;
4191                 if (nr < ARRAY_SIZE(pagepool))
4192                         continue;
4193                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4194                                    &bio, 0, &bio_flags, &prev_em_start);
4195                 nr = 0;
4196         }
4197         if (nr)
4198                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4199                                    &bio, 0, &bio_flags, &prev_em_start);
4200
4201         if (em_cached)
4202                 free_extent_map(em_cached);
4203
4204         BUG_ON(!list_empty(pages));
4205         if (bio)
4206                 return submit_one_bio(bio, 0, bio_flags);
4207         return 0;
4208 }
4209
4210 /*
4211  * basic invalidatepage code, this waits on any locked or writeback
4212  * ranges corresponding to the page, and then deletes any extent state
4213  * records from the tree
4214  */
4215 int extent_invalidatepage(struct extent_io_tree *tree,
4216                           struct page *page, unsigned long offset)
4217 {
4218         struct extent_state *cached_state = NULL;
4219         u64 start = page_offset(page);
4220         u64 end = start + PAGE_SIZE - 1;
4221         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4222
4223         start += ALIGN(offset, blocksize);
4224         if (start > end)
4225                 return 0;
4226
4227         lock_extent_bits(tree, start, end, &cached_state);
4228         wait_on_page_writeback(page);
4229         clear_extent_bit(tree, start, end,
4230                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4231                          EXTENT_DO_ACCOUNTING,
4232                          1, 1, &cached_state, GFP_NOFS);
4233         return 0;
4234 }
4235
4236 /*
4237  * a helper for releasepage, this tests for areas of the page that
4238  * are locked or under IO and drops the related state bits if it is safe
4239  * to drop the page.
4240  */
4241 static int try_release_extent_state(struct extent_map_tree *map,
4242                                     struct extent_io_tree *tree,
4243                                     struct page *page, gfp_t mask)
4244 {
4245         u64 start = page_offset(page);
4246         u64 end = start + PAGE_SIZE - 1;
4247         int ret = 1;
4248
4249         if (test_range_bit(tree, start, end,
4250                            EXTENT_IOBITS, 0, NULL))
4251                 ret = 0;
4252         else {
4253                 /*
4254                  * at this point we can safely clear everything except the
4255                  * locked bit and the nodatasum bit
4256                  */
4257                 ret = clear_extent_bit(tree, start, end,
4258                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4259                                  0, 0, NULL, mask);
4260
4261                 /* if clear_extent_bit failed for enomem reasons,
4262                  * we can't allow the release to continue.
4263                  */
4264                 if (ret < 0)
4265                         ret = 0;
4266                 else
4267                         ret = 1;
4268         }
4269         return ret;
4270 }
4271
4272 /*
4273  * a helper for releasepage.  As long as there are no locked extents
4274  * in the range corresponding to the page, both state records and extent
4275  * map records are removed
4276  */
4277 int try_release_extent_mapping(struct extent_map_tree *map,
4278                                struct extent_io_tree *tree, struct page *page,
4279                                gfp_t mask)
4280 {
4281         struct extent_map *em;
4282         u64 start = page_offset(page);
4283         u64 end = start + PAGE_SIZE - 1;
4284
4285         if (gfpflags_allow_blocking(mask) &&
4286             page->mapping->host->i_size > SZ_16M) {
4287                 u64 len;
4288                 while (start <= end) {
4289                         len = end - start + 1;
4290                         write_lock(&map->lock);
4291                         em = lookup_extent_mapping(map, start, len);
4292                         if (!em) {
4293                                 write_unlock(&map->lock);
4294                                 break;
4295                         }
4296                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4297                             em->start != start) {
4298                                 write_unlock(&map->lock);
4299                                 free_extent_map(em);
4300                                 break;
4301                         }
4302                         if (!test_range_bit(tree, em->start,
4303                                             extent_map_end(em) - 1,
4304                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4305                                             0, NULL)) {
4306                                 remove_extent_mapping(map, em);
4307                                 /* once for the rb tree */
4308                                 free_extent_map(em);
4309                         }
4310                         start = extent_map_end(em);
4311                         write_unlock(&map->lock);
4312
4313                         /* once for us */
4314                         free_extent_map(em);
4315                 }
4316         }
4317         return try_release_extent_state(map, tree, page, mask);
4318 }
4319
4320 /*
4321  * helper function for fiemap, which doesn't want to see any holes.
4322  * This maps until we find something past 'last'
4323  */
4324 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4325                                                 u64 offset,
4326                                                 u64 last,
4327                                                 get_extent_t *get_extent)
4328 {
4329         u64 sectorsize = btrfs_inode_sectorsize(inode);
4330         struct extent_map *em;
4331         u64 len;
4332
4333         if (offset >= last)
4334                 return NULL;
4335
4336         while (1) {
4337                 len = last - offset;
4338                 if (len == 0)
4339                         break;
4340                 len = ALIGN(len, sectorsize);
4341                 em = get_extent(BTRFS_I(inode), NULL, 0, offset, len, 0);
4342                 if (IS_ERR_OR_NULL(em))
4343                         return em;
4344
4345                 /* if this isn't a hole return it */
4346                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4347                     em->block_start != EXTENT_MAP_HOLE) {
4348                         return em;
4349                 }
4350
4351                 /* this is a hole, advance to the next extent */
4352                 offset = extent_map_end(em);
4353                 free_extent_map(em);
4354                 if (offset >= last)
4355                         break;
4356         }
4357         return NULL;
4358 }
4359
4360 /*
4361  * To cache previous fiemap extent
4362  *
4363  * Will be used for merging fiemap extent
4364  */
4365 struct fiemap_cache {
4366         u64 offset;
4367         u64 phys;
4368         u64 len;
4369         u32 flags;
4370         bool cached;
4371 };
4372
4373 /*
4374  * Helper to submit fiemap extent.
4375  *
4376  * Will try to merge current fiemap extent specified by @offset, @phys,
4377  * @len and @flags with cached one.
4378  * And only when we fails to merge, cached one will be submitted as
4379  * fiemap extent.
4380  *
4381  * Return value is the same as fiemap_fill_next_extent().
4382  */
4383 static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
4384                                 struct fiemap_cache *cache,
4385                                 u64 offset, u64 phys, u64 len, u32 flags)
4386 {
4387         int ret = 0;
4388
4389         if (!cache->cached)
4390                 goto assign;
4391
4392         /*
4393          * Sanity check, extent_fiemap() should have ensured that new
4394          * fiemap extent won't overlap with cahced one.
4395          * Not recoverable.
4396          *
4397          * NOTE: Physical address can overlap, due to compression
4398          */
4399         if (cache->offset + cache->len > offset) {
4400                 WARN_ON(1);
4401                 return -EINVAL;
4402         }
4403
4404         /*
4405          * Only merges fiemap extents if
4406          * 1) Their logical addresses are continuous
4407          *
4408          * 2) Their physical addresses are continuous
4409          *    So truly compressed (physical size smaller than logical size)
4410          *    extents won't get merged with each other
4411          *
4412          * 3) Share same flags except FIEMAP_EXTENT_LAST
4413          *    So regular extent won't get merged with prealloc extent
4414          */
4415         if (cache->offset + cache->len  == offset &&
4416             cache->phys + cache->len == phys  &&
4417             (cache->flags & ~FIEMAP_EXTENT_LAST) ==
4418                         (flags & ~FIEMAP_EXTENT_LAST)) {
4419                 cache->len += len;
4420                 cache->flags |= flags;
4421                 goto try_submit_last;
4422         }
4423
4424         /* Not mergeable, need to submit cached one */
4425         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4426                                       cache->len, cache->flags);
4427         cache->cached = false;
4428         if (ret)
4429                 return ret;
4430 assign:
4431         cache->cached = true;
4432         cache->offset = offset;
4433         cache->phys = phys;
4434         cache->len = len;
4435         cache->flags = flags;
4436 try_submit_last:
4437         if (cache->flags & FIEMAP_EXTENT_LAST) {
4438                 ret = fiemap_fill_next_extent(fieinfo, cache->offset,
4439                                 cache->phys, cache->len, cache->flags);
4440                 cache->cached = false;
4441         }
4442         return ret;
4443 }
4444
4445 /*
4446  * Emit last fiemap cache
4447  *
4448  * The last fiemap cache may still be cached in the following case:
4449  * 0                  4k                    8k
4450  * |<- Fiemap range ->|
4451  * |<------------  First extent ----------->|
4452  *
4453  * In this case, the first extent range will be cached but not emitted.
4454  * So we must emit it before ending extent_fiemap().
4455  */
4456 static int emit_last_fiemap_cache(struct btrfs_fs_info *fs_info,
4457                                   struct fiemap_extent_info *fieinfo,
4458                                   struct fiemap_cache *cache)
4459 {
4460         int ret;
4461
4462         if (!cache->cached)
4463                 return 0;
4464
4465         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4466                                       cache->len, cache->flags);
4467         cache->cached = false;
4468         if (ret > 0)
4469                 ret = 0;
4470         return ret;
4471 }
4472
4473 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4474                 __u64 start, __u64 len, get_extent_t *get_extent)
4475 {
4476         int ret = 0;
4477         u64 off = start;
4478         u64 max = start + len;
4479         u32 flags = 0;
4480         u32 found_type;
4481         u64 last;
4482         u64 last_for_get_extent = 0;
4483         u64 disko = 0;
4484         u64 isize = i_size_read(inode);
4485         struct btrfs_key found_key;
4486         struct extent_map *em = NULL;
4487         struct extent_state *cached_state = NULL;
4488         struct btrfs_path *path;
4489         struct btrfs_root *root = BTRFS_I(inode)->root;
4490         struct fiemap_cache cache = { 0 };
4491         int end = 0;
4492         u64 em_start = 0;
4493         u64 em_len = 0;
4494         u64 em_end = 0;
4495
4496         if (len == 0)
4497                 return -EINVAL;
4498
4499         path = btrfs_alloc_path();
4500         if (!path)
4501                 return -ENOMEM;
4502         path->leave_spinning = 1;
4503
4504         start = round_down(start, btrfs_inode_sectorsize(inode));
4505         len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
4506
4507         /*
4508          * lookup the last file extent.  We're not using i_size here
4509          * because there might be preallocation past i_size
4510          */
4511         ret = btrfs_lookup_file_extent(NULL, root, path,
4512                         btrfs_ino(BTRFS_I(inode)), -1, 0);
4513         if (ret < 0) {
4514                 btrfs_free_path(path);
4515                 return ret;
4516         } else {
4517                 WARN_ON(!ret);
4518                 if (ret == 1)
4519                         ret = 0;
4520         }
4521
4522         path->slots[0]--;
4523         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4524         found_type = found_key.type;
4525
4526         /* No extents, but there might be delalloc bits */
4527         if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) ||
4528             found_type != BTRFS_EXTENT_DATA_KEY) {
4529                 /* have to trust i_size as the end */
4530                 last = (u64)-1;
4531                 last_for_get_extent = isize;
4532         } else {
4533                 /*
4534                  * remember the start of the last extent.  There are a
4535                  * bunch of different factors that go into the length of the
4536                  * extent, so its much less complex to remember where it started
4537                  */
4538                 last = found_key.offset;
4539                 last_for_get_extent = last + 1;
4540         }
4541         btrfs_release_path(path);
4542
4543         /*
4544          * we might have some extents allocated but more delalloc past those
4545          * extents.  so, we trust isize unless the start of the last extent is
4546          * beyond isize
4547          */
4548         if (last < isize) {
4549                 last = (u64)-1;
4550                 last_for_get_extent = isize;
4551         }
4552
4553         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4554                          &cached_state);
4555
4556         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4557                                    get_extent);
4558         if (!em)
4559                 goto out;
4560         if (IS_ERR(em)) {
4561                 ret = PTR_ERR(em);
4562                 goto out;
4563         }
4564
4565         while (!end) {
4566                 u64 offset_in_extent = 0;
4567
4568                 /* break if the extent we found is outside the range */
4569                 if (em->start >= max || extent_map_end(em) < off)
4570                         break;
4571
4572                 /*
4573                  * get_extent may return an extent that starts before our
4574                  * requested range.  We have to make sure the ranges
4575                  * we return to fiemap always move forward and don't
4576                  * overlap, so adjust the offsets here
4577                  */
4578                 em_start = max(em->start, off);
4579
4580                 /*
4581                  * record the offset from the start of the extent
4582                  * for adjusting the disk offset below.  Only do this if the
4583                  * extent isn't compressed since our in ram offset may be past
4584                  * what we have actually allocated on disk.
4585                  */
4586                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4587                         offset_in_extent = em_start - em->start;
4588                 em_end = extent_map_end(em);
4589                 em_len = em_end - em_start;
4590                 disko = 0;
4591                 flags = 0;
4592
4593                 /*
4594                  * bump off for our next call to get_extent
4595                  */
4596                 off = extent_map_end(em);
4597                 if (off >= max)
4598                         end = 1;
4599
4600                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4601                         end = 1;
4602                         flags |= FIEMAP_EXTENT_LAST;
4603                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4604                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4605                                   FIEMAP_EXTENT_NOT_ALIGNED);
4606                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4607                         flags |= (FIEMAP_EXTENT_DELALLOC |
4608                                   FIEMAP_EXTENT_UNKNOWN);
4609                 } else if (fieinfo->fi_extents_max) {
4610                         struct btrfs_trans_handle *trans;
4611
4612                         u64 bytenr = em->block_start -
4613                                 (em->start - em->orig_start);
4614
4615                         disko = em->block_start + offset_in_extent;
4616
4617                         /*
4618                          * We need a trans handle to get delayed refs
4619                          */
4620                         trans = btrfs_join_transaction(root);
4621                         /*
4622                          * It's OK if we can't start a trans we can still check
4623                          * from commit_root
4624                          */
4625                         if (IS_ERR(trans))
4626                                 trans = NULL;
4627
4628                         /*
4629                          * As btrfs supports shared space, this information
4630                          * can be exported to userspace tools via
4631                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4632                          * then we're just getting a count and we can skip the
4633                          * lookup stuff.
4634                          */
4635                         ret = btrfs_check_shared(trans, root->fs_info,
4636                                         root->objectid,
4637                                         btrfs_ino(BTRFS_I(inode)), bytenr);
4638                         if (trans)
4639                                 btrfs_end_transaction(trans);
4640                         if (ret < 0)
4641                                 goto out_free;
4642                         if (ret)
4643                                 flags |= FIEMAP_EXTENT_SHARED;
4644                         ret = 0;
4645                 }
4646                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4647                         flags |= FIEMAP_EXTENT_ENCODED;
4648                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4649                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4650
4651                 free_extent_map(em);
4652                 em = NULL;
4653                 if ((em_start >= last) || em_len == (u64)-1 ||
4654                    (last == (u64)-1 && isize <= em_end)) {
4655                         flags |= FIEMAP_EXTENT_LAST;
4656                         end = 1;
4657                 }
4658
4659                 /* now scan forward to see if this is really the last extent. */
4660                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4661                                            get_extent);
4662                 if (IS_ERR(em)) {
4663                         ret = PTR_ERR(em);
4664                         goto out;
4665                 }
4666                 if (!em) {
4667                         flags |= FIEMAP_EXTENT_LAST;
4668                         end = 1;
4669                 }
4670                 ret = emit_fiemap_extent(fieinfo, &cache, em_start, disko,
4671                                            em_len, flags);
4672                 if (ret) {
4673                         if (ret == 1)
4674                                 ret = 0;
4675                         goto out_free;
4676                 }
4677         }
4678 out_free:
4679         if (!ret)
4680                 ret = emit_last_fiemap_cache(root->fs_info, fieinfo, &cache);
4681         free_extent_map(em);
4682 out:
4683         btrfs_free_path(path);
4684         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4685                              &cached_state, GFP_NOFS);
4686         return ret;
4687 }
4688
4689 static void __free_extent_buffer(struct extent_buffer *eb)
4690 {
4691         btrfs_leak_debug_del(&eb->leak_list);
4692         kmem_cache_free(extent_buffer_cache, eb);
4693 }
4694
4695 int extent_buffer_under_io(struct extent_buffer *eb)
4696 {
4697         return (atomic_read(&eb->io_pages) ||
4698                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4699                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4700 }
4701
4702 /*
4703  * Helper for releasing extent buffer page.
4704  */
4705 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4706 {
4707         unsigned long index;
4708         struct page *page;
4709         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4710
4711         BUG_ON(extent_buffer_under_io(eb));
4712
4713         index = num_extent_pages(eb->start, eb->len);
4714         if (index == 0)
4715                 return;
4716
4717         do {
4718                 index--;
4719                 page = eb->pages[index];
4720                 if (!page)
4721                         continue;
4722                 if (mapped)
4723                         spin_lock(&page->mapping->private_lock);
4724                 /*
4725                  * We do this since we'll remove the pages after we've
4726                  * removed the eb from the radix tree, so we could race
4727                  * and have this page now attached to the new eb.  So
4728                  * only clear page_private if it's still connected to
4729                  * this eb.
4730                  */
4731                 if (PagePrivate(page) &&
4732                     page->private == (unsigned long)eb) {
4733                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4734                         BUG_ON(PageDirty(page));
4735                         BUG_ON(PageWriteback(page));
4736                         /*
4737                          * We need to make sure we haven't be attached
4738                          * to a new eb.
4739                          */
4740                         ClearPagePrivate(page);
4741                         set_page_private(page, 0);
4742                         /* One for the page private */
4743                         put_page(page);
4744                 }
4745
4746                 if (mapped)
4747                         spin_unlock(&page->mapping->private_lock);
4748
4749                 /* One for when we allocated the page */
4750                 put_page(page);
4751         } while (index != 0);
4752 }
4753
4754 /*
4755  * Helper for releasing the extent buffer.
4756  */
4757 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4758 {
4759         btrfs_release_extent_buffer_page(eb);
4760         __free_extent_buffer(eb);
4761 }
4762
4763 static struct extent_buffer *
4764 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4765                       unsigned long len)
4766 {
4767         struct extent_buffer *eb = NULL;
4768
4769         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4770         eb->start = start;
4771         eb->len = len;
4772         eb->fs_info = fs_info;
4773         eb->bflags = 0;
4774         rwlock_init(&eb->lock);
4775         atomic_set(&eb->write_locks, 0);
4776         atomic_set(&eb->read_locks, 0);
4777         atomic_set(&eb->blocking_readers, 0);
4778         atomic_set(&eb->blocking_writers, 0);
4779         atomic_set(&eb->spinning_readers, 0);
4780         atomic_set(&eb->spinning_writers, 0);
4781         eb->lock_nested = 0;
4782         init_waitqueue_head(&eb->write_lock_wq);
4783         init_waitqueue_head(&eb->read_lock_wq);
4784
4785         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4786
4787         spin_lock_init(&eb->refs_lock);
4788         atomic_set(&eb->refs, 1);
4789         atomic_set(&eb->io_pages, 0);
4790
4791         /*
4792          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4793          */
4794         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4795                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4796         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4797
4798         return eb;
4799 }
4800
4801 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4802 {
4803         unsigned long i;
4804         struct page *p;
4805         struct extent_buffer *new;
4806         unsigned long num_pages = num_extent_pages(src->start, src->len);
4807
4808         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4809         if (new == NULL)
4810                 return NULL;
4811
4812         for (i = 0; i < num_pages; i++) {
4813                 p = alloc_page(GFP_NOFS);
4814                 if (!p) {
4815                         btrfs_release_extent_buffer(new);
4816                         return NULL;
4817                 }
4818                 attach_extent_buffer_page(new, p);
4819                 WARN_ON(PageDirty(p));
4820                 SetPageUptodate(p);
4821                 new->pages[i] = p;
4822                 copy_page(page_address(p), page_address(src->pages[i]));
4823         }
4824
4825         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4826         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4827
4828         return new;
4829 }
4830
4831 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4832                                                   u64 start, unsigned long len)
4833 {
4834         struct extent_buffer *eb;
4835         unsigned long num_pages;
4836         unsigned long i;
4837
4838         num_pages = num_extent_pages(start, len);
4839
4840         eb = __alloc_extent_buffer(fs_info, start, len);
4841         if (!eb)
4842                 return NULL;
4843
4844         for (i = 0; i < num_pages; i++) {
4845                 eb->pages[i] = alloc_page(GFP_NOFS);
4846                 if (!eb->pages[i])
4847                         goto err;
4848         }
4849         set_extent_buffer_uptodate(eb);
4850         btrfs_set_header_nritems(eb, 0);
4851         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4852
4853         return eb;
4854 err:
4855         for (; i > 0; i--)
4856                 __free_page(eb->pages[i - 1]);
4857         __free_extent_buffer(eb);
4858         return NULL;
4859 }
4860
4861 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4862                                                 u64 start)
4863 {
4864         return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
4865 }
4866
4867 static void check_buffer_tree_ref(struct extent_buffer *eb)
4868 {
4869         int refs;
4870         /* the ref bit is tricky.  We have to make sure it is set
4871          * if we have the buffer dirty.   Otherwise the
4872          * code to free a buffer can end up dropping a dirty
4873          * page
4874          *
4875          * Once the ref bit is set, it won't go away while the
4876          * buffer is dirty or in writeback, and it also won't
4877          * go away while we have the reference count on the
4878          * eb bumped.
4879          *
4880          * We can't just set the ref bit without bumping the
4881          * ref on the eb because free_extent_buffer might
4882          * see the ref bit and try to clear it.  If this happens
4883          * free_extent_buffer might end up dropping our original
4884          * ref by mistake and freeing the page before we are able
4885          * to add one more ref.
4886          *
4887          * So bump the ref count first, then set the bit.  If someone
4888          * beat us to it, drop the ref we added.
4889          */
4890         refs = atomic_read(&eb->refs);
4891         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4892                 return;
4893
4894         spin_lock(&eb->refs_lock);
4895         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4896                 atomic_inc(&eb->refs);
4897         spin_unlock(&eb->refs_lock);
4898 }
4899
4900 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4901                 struct page *accessed)
4902 {
4903         unsigned long num_pages, i;
4904
4905         check_buffer_tree_ref(eb);
4906
4907         num_pages = num_extent_pages(eb->start, eb->len);
4908         for (i = 0; i < num_pages; i++) {
4909                 struct page *p = eb->pages[i];
4910
4911                 if (p != accessed)
4912                         mark_page_accessed(p);
4913         }
4914 }
4915
4916 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4917                                          u64 start)
4918 {
4919         struct extent_buffer *eb;
4920
4921         rcu_read_lock();
4922         eb = radix_tree_lookup(&fs_info->buffer_radix,
4923                                start >> PAGE_SHIFT);
4924         if (eb && atomic_inc_not_zero(&eb->refs)) {
4925                 rcu_read_unlock();
4926                 /*
4927                  * Lock our eb's refs_lock to avoid races with
4928                  * free_extent_buffer. When we get our eb it might be flagged
4929                  * with EXTENT_BUFFER_STALE and another task running
4930                  * free_extent_buffer might have seen that flag set,
4931                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4932                  * writeback flags not set) and it's still in the tree (flag
4933                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4934                  * of decrementing the extent buffer's reference count twice.
4935                  * So here we could race and increment the eb's reference count,
4936                  * clear its stale flag, mark it as dirty and drop our reference
4937                  * before the other task finishes executing free_extent_buffer,
4938                  * which would later result in an attempt to free an extent
4939                  * buffer that is dirty.
4940                  */
4941                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4942                         spin_lock(&eb->refs_lock);
4943                         spin_unlock(&eb->refs_lock);
4944                 }
4945                 mark_extent_buffer_accessed(eb, NULL);
4946                 return eb;
4947         }
4948         rcu_read_unlock();
4949
4950         return NULL;
4951 }
4952
4953 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4954 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4955                                         u64 start)
4956 {
4957         struct extent_buffer *eb, *exists = NULL;
4958         int ret;
4959
4960         eb = find_extent_buffer(fs_info, start);
4961         if (eb)
4962                 return eb;
4963         eb = alloc_dummy_extent_buffer(fs_info, start);
4964         if (!eb)
4965                 return NULL;
4966         eb->fs_info = fs_info;
4967 again:
4968         ret = radix_tree_preload(GFP_NOFS);
4969         if (ret)
4970                 goto free_eb;
4971         spin_lock(&fs_info->buffer_lock);
4972         ret = radix_tree_insert(&fs_info->buffer_radix,
4973                                 start >> PAGE_SHIFT, eb);
4974         spin_unlock(&fs_info->buffer_lock);
4975         radix_tree_preload_end();
4976         if (ret == -EEXIST) {
4977                 exists = find_extent_buffer(fs_info, start);
4978                 if (exists)
4979                         goto free_eb;
4980                 else
4981                         goto again;
4982         }
4983         check_buffer_tree_ref(eb);
4984         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4985
4986         /*
4987          * We will free dummy extent buffer's if they come into
4988          * free_extent_buffer with a ref count of 2, but if we are using this we
4989          * want the buffers to stay in memory until we're done with them, so
4990          * bump the ref count again.
4991          */
4992         atomic_inc(&eb->refs);
4993         return eb;
4994 free_eb:
4995         btrfs_release_extent_buffer(eb);
4996         return exists;
4997 }
4998 #endif
4999
5000 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
5001                                           u64 start)
5002 {
5003         unsigned long len = fs_info->nodesize;
5004         unsigned long num_pages = num_extent_pages(start, len);
5005         unsigned long i;
5006         unsigned long index = start >> PAGE_SHIFT;
5007         struct extent_buffer *eb;
5008         struct extent_buffer *exists = NULL;
5009         struct page *p;
5010         struct address_space *mapping = fs_info->btree_inode->i_mapping;
5011         int uptodate = 1;
5012         int ret;
5013
5014         if (!IS_ALIGNED(start, fs_info->sectorsize)) {
5015                 btrfs_err(fs_info, "bad tree block start %llu", start);
5016                 return ERR_PTR(-EINVAL);
5017         }
5018
5019         eb = find_extent_buffer(fs_info, start);
5020         if (eb)
5021                 return eb;
5022
5023         eb = __alloc_extent_buffer(fs_info, start, len);
5024         if (!eb)
5025                 return ERR_PTR(-ENOMEM);
5026
5027         for (i = 0; i < num_pages; i++, index++) {
5028                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
5029                 if (!p) {
5030                         exists = ERR_PTR(-ENOMEM);
5031                         goto free_eb;
5032                 }
5033
5034                 spin_lock(&mapping->private_lock);
5035                 if (PagePrivate(p)) {
5036                         /*
5037                          * We could have already allocated an eb for this page
5038                          * and attached one so lets see if we can get a ref on
5039                          * the existing eb, and if we can we know it's good and
5040                          * we can just return that one, else we know we can just
5041                          * overwrite page->private.
5042                          */
5043                         exists = (struct extent_buffer *)p->private;
5044                         if (atomic_inc_not_zero(&exists->refs)) {
5045                                 spin_unlock(&mapping->private_lock);
5046                                 unlock_page(p);
5047                                 put_page(p);
5048                                 mark_extent_buffer_accessed(exists, p);
5049                                 goto free_eb;
5050                         }
5051                         exists = NULL;
5052
5053                         /*
5054                          * Do this so attach doesn't complain and we need to
5055                          * drop the ref the old guy had.
5056                          */
5057                         ClearPagePrivate(p);
5058                         WARN_ON(PageDirty(p));
5059                         put_page(p);
5060                 }
5061                 attach_extent_buffer_page(eb, p);
5062                 spin_unlock(&mapping->private_lock);
5063                 WARN_ON(PageDirty(p));
5064                 eb->pages[i] = p;
5065                 if (!PageUptodate(p))
5066                         uptodate = 0;
5067
5068                 /*
5069                  * see below about how we avoid a nasty race with release page
5070                  * and why we unlock later
5071                  */
5072         }
5073         if (uptodate)
5074                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5075 again:
5076         ret = radix_tree_preload(GFP_NOFS);
5077         if (ret) {
5078                 exists = ERR_PTR(ret);
5079                 goto free_eb;
5080         }
5081
5082         spin_lock(&fs_info->buffer_lock);
5083         ret = radix_tree_insert(&fs_info->buffer_radix,
5084                                 start >> PAGE_SHIFT, eb);
5085         spin_unlock(&fs_info->buffer_lock);
5086         radix_tree_preload_end();
5087         if (ret == -EEXIST) {
5088                 exists = find_extent_buffer(fs_info, start);
5089                 if (exists)
5090                         goto free_eb;
5091                 else
5092                         goto again;
5093         }
5094         /* add one reference for the tree */
5095         check_buffer_tree_ref(eb);
5096         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5097
5098         /*
5099          * there is a race where release page may have
5100          * tried to find this extent buffer in the radix
5101          * but failed.  It will tell the VM it is safe to
5102          * reclaim the, and it will clear the page private bit.
5103          * We must make sure to set the page private bit properly
5104          * after the extent buffer is in the radix tree so
5105          * it doesn't get lost
5106          */
5107         SetPageChecked(eb->pages[0]);
5108         for (i = 1; i < num_pages; i++) {
5109                 p = eb->pages[i];
5110                 ClearPageChecked(p);
5111                 unlock_page(p);
5112         }
5113         unlock_page(eb->pages[0]);
5114         return eb;
5115
5116 free_eb:
5117         WARN_ON(!atomic_dec_and_test(&eb->refs));
5118         for (i = 0; i < num_pages; i++) {
5119                 if (eb->pages[i])
5120                         unlock_page(eb->pages[i]);
5121         }
5122
5123         btrfs_release_extent_buffer(eb);
5124         return exists;
5125 }
5126
5127 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5128 {
5129         struct extent_buffer *eb =
5130                         container_of(head, struct extent_buffer, rcu_head);
5131
5132         __free_extent_buffer(eb);
5133 }
5134
5135 /* Expects to have eb->eb_lock already held */
5136 static int release_extent_buffer(struct extent_buffer *eb)
5137 {
5138         WARN_ON(atomic_read(&eb->refs) == 0);
5139         if (atomic_dec_and_test(&eb->refs)) {
5140                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5141                         struct btrfs_fs_info *fs_info = eb->fs_info;
5142
5143                         spin_unlock(&eb->refs_lock);
5144
5145                         spin_lock(&fs_info->buffer_lock);
5146                         radix_tree_delete(&fs_info->buffer_radix,
5147                                           eb->start >> PAGE_SHIFT);
5148                         spin_unlock(&fs_info->buffer_lock);
5149                 } else {
5150                         spin_unlock(&eb->refs_lock);
5151                 }
5152
5153                 /* Should be safe to release our pages at this point */
5154                 btrfs_release_extent_buffer_page(eb);
5155 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5156                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5157                         __free_extent_buffer(eb);
5158                         return 1;
5159                 }
5160 #endif
5161                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5162                 return 1;
5163         }
5164         spin_unlock(&eb->refs_lock);
5165
5166         return 0;
5167 }
5168
5169 void free_extent_buffer(struct extent_buffer *eb)
5170 {
5171         int refs;
5172         int old;
5173         if (!eb)
5174                 return;
5175
5176         while (1) {
5177                 refs = atomic_read(&eb->refs);
5178                 if (refs <= 3)
5179                         break;
5180                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5181                 if (old == refs)
5182                         return;
5183         }
5184
5185         spin_lock(&eb->refs_lock);
5186         if (atomic_read(&eb->refs) == 2 &&
5187             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5188                 atomic_dec(&eb->refs);
5189
5190         if (atomic_read(&eb->refs) == 2 &&
5191             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5192             !extent_buffer_under_io(eb) &&
5193             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5194                 atomic_dec(&eb->refs);
5195
5196         /*
5197          * I know this is terrible, but it's temporary until we stop tracking
5198          * the uptodate bits and such for the extent buffers.
5199          */
5200         release_extent_buffer(eb);
5201 }
5202
5203 void free_extent_buffer_stale(struct extent_buffer *eb)
5204 {
5205         if (!eb)
5206                 return;
5207
5208         spin_lock(&eb->refs_lock);
5209         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5210
5211         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5212             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5213                 atomic_dec(&eb->refs);
5214         release_extent_buffer(eb);
5215 }
5216
5217 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5218 {
5219         unsigned long i;
5220         unsigned long num_pages;
5221         struct page *page;
5222
5223         num_pages = num_extent_pages(eb->start, eb->len);
5224
5225         for (i = 0; i < num_pages; i++) {
5226                 page = eb->pages[i];
5227                 if (!PageDirty(page))
5228                         continue;
5229
5230                 lock_page(page);
5231                 WARN_ON(!PagePrivate(page));
5232
5233                 clear_page_dirty_for_io(page);
5234                 spin_lock_irq(&page->mapping->tree_lock);
5235                 if (!PageDirty(page)) {
5236                         radix_tree_tag_clear(&page->mapping->page_tree,
5237                                                 page_index(page),
5238                                                 PAGECACHE_TAG_DIRTY);
5239                 }
5240                 spin_unlock_irq(&page->mapping->tree_lock);
5241                 ClearPageError(page);
5242                 unlock_page(page);
5243         }
5244         WARN_ON(atomic_read(&eb->refs) == 0);
5245 }
5246
5247 int set_extent_buffer_dirty(struct extent_buffer *eb)
5248 {
5249         unsigned long i;
5250         unsigned long num_pages;
5251         int was_dirty = 0;
5252
5253         check_buffer_tree_ref(eb);
5254
5255         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5256
5257         num_pages = num_extent_pages(eb->start, eb->len);
5258         WARN_ON(atomic_read(&eb->refs) == 0);
5259         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5260
5261         for (i = 0; i < num_pages; i++)
5262                 set_page_dirty(eb->pages[i]);
5263         return was_dirty;
5264 }
5265
5266 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5267 {
5268         unsigned long i;
5269         struct page *page;
5270         unsigned long num_pages;
5271
5272         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5273         num_pages = num_extent_pages(eb->start, eb->len);
5274         for (i = 0; i < num_pages; i++) {
5275                 page = eb->pages[i];
5276                 if (page)
5277                         ClearPageUptodate(page);
5278         }
5279 }
5280
5281 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5282 {
5283         unsigned long i;
5284         struct page *page;
5285         unsigned long num_pages;
5286
5287         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5288         num_pages = num_extent_pages(eb->start, eb->len);
5289         for (i = 0; i < num_pages; i++) {
5290                 page = eb->pages[i];
5291                 SetPageUptodate(page);
5292         }
5293 }
5294
5295 int extent_buffer_uptodate(struct extent_buffer *eb)
5296 {
5297         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5298 }
5299
5300 int read_extent_buffer_pages(struct extent_io_tree *tree,
5301                              struct extent_buffer *eb, int wait,
5302                              get_extent_t *get_extent, int mirror_num)
5303 {
5304         unsigned long i;
5305         struct page *page;
5306         int err;
5307         int ret = 0;
5308         int locked_pages = 0;
5309         int all_uptodate = 1;
5310         unsigned long num_pages;
5311         unsigned long num_reads = 0;
5312         struct bio *bio = NULL;
5313         unsigned long bio_flags = 0;
5314
5315         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5316                 return 0;
5317
5318         num_pages = num_extent_pages(eb->start, eb->len);
5319         for (i = 0; i < num_pages; i++) {
5320                 page = eb->pages[i];
5321                 if (wait == WAIT_NONE) {
5322                         if (!trylock_page(page))
5323                                 goto unlock_exit;
5324                 } else {
5325                         lock_page(page);
5326                 }
5327                 locked_pages++;
5328         }
5329         /*
5330          * We need to firstly lock all pages to make sure that
5331          * the uptodate bit of our pages won't be affected by
5332          * clear_extent_buffer_uptodate().
5333          */
5334         for (i = 0; i < num_pages; i++) {
5335                 page = eb->pages[i];
5336                 if (!PageUptodate(page)) {
5337                         num_reads++;
5338                         all_uptodate = 0;
5339                 }
5340         }
5341
5342         if (all_uptodate) {
5343                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5344                 goto unlock_exit;
5345         }
5346
5347         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5348         eb->read_mirror = 0;
5349         atomic_set(&eb->io_pages, num_reads);
5350         for (i = 0; i < num_pages; i++) {
5351                 page = eb->pages[i];
5352
5353                 if (!PageUptodate(page)) {
5354                         if (ret) {
5355                                 atomic_dec(&eb->io_pages);
5356                                 unlock_page(page);
5357                                 continue;
5358                         }
5359
5360                         ClearPageError(page);
5361                         err = __extent_read_full_page(tree, page,
5362                                                       get_extent, &bio,
5363                                                       mirror_num, &bio_flags,
5364                                                       REQ_META);
5365                         if (err) {
5366                                 ret = err;
5367                                 /*
5368                                  * We use &bio in above __extent_read_full_page,
5369                                  * so we ensure that if it returns error, the
5370                                  * current page fails to add itself to bio and
5371                                  * it's been unlocked.
5372                                  *
5373                                  * We must dec io_pages by ourselves.
5374                                  */
5375                                 atomic_dec(&eb->io_pages);
5376                         }
5377                 } else {
5378                         unlock_page(page);
5379                 }
5380         }
5381
5382         if (bio) {
5383                 err = submit_one_bio(bio, mirror_num, bio_flags);
5384                 if (err)
5385                         return err;
5386         }
5387
5388         if (ret || wait != WAIT_COMPLETE)
5389                 return ret;
5390
5391         for (i = 0; i < num_pages; i++) {
5392                 page = eb->pages[i];
5393                 wait_on_page_locked(page);
5394                 if (!PageUptodate(page))
5395                         ret = -EIO;
5396         }
5397
5398         return ret;
5399
5400 unlock_exit:
5401         while (locked_pages > 0) {
5402                 locked_pages--;
5403                 page = eb->pages[locked_pages];
5404                 unlock_page(page);
5405         }
5406         return ret;
5407 }
5408
5409 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5410                         unsigned long start,
5411                         unsigned long len)
5412 {
5413         size_t cur;
5414         size_t offset;
5415         struct page *page;
5416         char *kaddr;
5417         char *dst = (char *)dstv;
5418         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5419         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5420
5421         WARN_ON(start > eb->len);
5422         WARN_ON(start + len > eb->start + eb->len);
5423
5424         offset = (start_offset + start) & (PAGE_SIZE - 1);
5425
5426         while (len > 0) {
5427                 page = eb->pages[i];
5428
5429                 cur = min(len, (PAGE_SIZE - offset));
5430                 kaddr = page_address(page);
5431                 memcpy(dst, kaddr + offset, cur);
5432
5433                 dst += cur;
5434                 len -= cur;
5435                 offset = 0;
5436                 i++;
5437         }
5438 }
5439
5440 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5441                         unsigned long start,
5442                         unsigned long len)
5443 {
5444         size_t cur;
5445         size_t offset;
5446         struct page *page;
5447         char *kaddr;
5448         char __user *dst = (char __user *)dstv;
5449         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5450         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5451         int ret = 0;
5452
5453         WARN_ON(start > eb->len);
5454         WARN_ON(start + len > eb->start + eb->len);
5455
5456         offset = (start_offset + start) & (PAGE_SIZE - 1);
5457
5458         while (len > 0) {
5459                 page = eb->pages[i];
5460
5461                 cur = min(len, (PAGE_SIZE - offset));
5462                 kaddr = page_address(page);
5463                 if (copy_to_user(dst, kaddr + offset, cur)) {
5464                         ret = -EFAULT;
5465                         break;
5466                 }
5467
5468                 dst += cur;
5469                 len -= cur;
5470                 offset = 0;
5471                 i++;
5472         }
5473
5474         return ret;
5475 }
5476
5477 /*
5478  * return 0 if the item is found within a page.
5479  * return 1 if the item spans two pages.
5480  * return -EINVAL otherwise.
5481  */
5482 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5483                                unsigned long min_len, char **map,
5484                                unsigned long *map_start,
5485                                unsigned long *map_len)
5486 {
5487         size_t offset = start & (PAGE_SIZE - 1);
5488         char *kaddr;
5489         struct page *p;
5490         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5491         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5492         unsigned long end_i = (start_offset + start + min_len - 1) >>
5493                 PAGE_SHIFT;
5494
5495         if (i != end_i)
5496                 return 1;
5497
5498         if (i == 0) {
5499                 offset = start_offset;
5500                 *map_start = 0;
5501         } else {
5502                 offset = 0;
5503                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5504         }
5505
5506         if (start + min_len > eb->len) {
5507                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5508                        eb->start, eb->len, start, min_len);
5509                 return -EINVAL;
5510         }
5511
5512         p = eb->pages[i];
5513         kaddr = page_address(p);
5514         *map = kaddr + offset;
5515         *map_len = PAGE_SIZE - offset;
5516         return 0;
5517 }
5518
5519 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5520                           unsigned long start,
5521                           unsigned long len)
5522 {
5523         size_t cur;
5524         size_t offset;
5525         struct page *page;
5526         char *kaddr;
5527         char *ptr = (char *)ptrv;
5528         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5529         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5530         int ret = 0;
5531
5532         WARN_ON(start > eb->len);
5533         WARN_ON(start + len > eb->start + eb->len);
5534
5535         offset = (start_offset + start) & (PAGE_SIZE - 1);
5536
5537         while (len > 0) {
5538                 page = eb->pages[i];
5539
5540                 cur = min(len, (PAGE_SIZE - offset));
5541
5542                 kaddr = page_address(page);
5543                 ret = memcmp(ptr, kaddr + offset, cur);
5544                 if (ret)
5545                         break;
5546
5547                 ptr += cur;
5548                 len -= cur;
5549                 offset = 0;
5550                 i++;
5551         }
5552         return ret;
5553 }
5554
5555 void write_extent_buffer_chunk_tree_uuid(struct extent_buffer *eb,
5556                 const void *srcv)
5557 {
5558         char *kaddr;
5559
5560         WARN_ON(!PageUptodate(eb->pages[0]));
5561         kaddr = page_address(eb->pages[0]);
5562         memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv,
5563                         BTRFS_FSID_SIZE);
5564 }
5565
5566 void write_extent_buffer_fsid(struct extent_buffer *eb, const void *srcv)
5567 {
5568         char *kaddr;
5569
5570         WARN_ON(!PageUptodate(eb->pages[0]));
5571         kaddr = page_address(eb->pages[0]);
5572         memcpy(kaddr + offsetof(struct btrfs_header, fsid), srcv,
5573                         BTRFS_FSID_SIZE);
5574 }
5575
5576 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5577                          unsigned long start, unsigned long len)
5578 {
5579         size_t cur;
5580         size_t offset;
5581         struct page *page;
5582         char *kaddr;
5583         char *src = (char *)srcv;
5584         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5585         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5586
5587         WARN_ON(start > eb->len);
5588         WARN_ON(start + len > eb->start + eb->len);
5589
5590         offset = (start_offset + start) & (PAGE_SIZE - 1);
5591
5592         while (len > 0) {
5593                 page = eb->pages[i];
5594                 WARN_ON(!PageUptodate(page));
5595
5596                 cur = min(len, PAGE_SIZE - offset);
5597                 kaddr = page_address(page);
5598                 memcpy(kaddr + offset, src, cur);
5599
5600                 src += cur;
5601                 len -= cur;
5602                 offset = 0;
5603                 i++;
5604         }
5605 }
5606
5607 void memzero_extent_buffer(struct extent_buffer *eb, unsigned long start,
5608                 unsigned long len)
5609 {
5610         size_t cur;
5611         size_t offset;
5612         struct page *page;
5613         char *kaddr;
5614         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5615         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5616
5617         WARN_ON(start > eb->len);
5618         WARN_ON(start + len > eb->start + eb->len);
5619
5620         offset = (start_offset + start) & (PAGE_SIZE - 1);
5621
5622         while (len > 0) {
5623                 page = eb->pages[i];
5624                 WARN_ON(!PageUptodate(page));
5625
5626                 cur = min(len, PAGE_SIZE - offset);
5627                 kaddr = page_address(page);
5628                 memset(kaddr + offset, 0, cur);
5629
5630                 len -= cur;
5631                 offset = 0;
5632                 i++;
5633         }
5634 }
5635
5636 void copy_extent_buffer_full(struct extent_buffer *dst,
5637                              struct extent_buffer *src)
5638 {
5639         int i;
5640         unsigned num_pages;
5641
5642         ASSERT(dst->len == src->len);
5643
5644         num_pages = num_extent_pages(dst->start, dst->len);
5645         for (i = 0; i < num_pages; i++)
5646                 copy_page(page_address(dst->pages[i]),
5647                                 page_address(src->pages[i]));
5648 }
5649
5650 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5651                         unsigned long dst_offset, unsigned long src_offset,
5652                         unsigned long len)
5653 {
5654         u64 dst_len = dst->len;
5655         size_t cur;
5656         size_t offset;
5657         struct page *page;
5658         char *kaddr;
5659         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5660         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5661
5662         WARN_ON(src->len != dst_len);
5663
5664         offset = (start_offset + dst_offset) &
5665                 (PAGE_SIZE - 1);
5666
5667         while (len > 0) {
5668                 page = dst->pages[i];
5669                 WARN_ON(!PageUptodate(page));
5670
5671                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5672
5673                 kaddr = page_address(page);
5674                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5675
5676                 src_offset += cur;
5677                 len -= cur;
5678                 offset = 0;
5679                 i++;
5680         }
5681 }
5682
5683 void le_bitmap_set(u8 *map, unsigned int start, int len)
5684 {
5685         u8 *p = map + BIT_BYTE(start);
5686         const unsigned int size = start + len;
5687         int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5688         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);
5689
5690         while (len - bits_to_set >= 0) {
5691                 *p |= mask_to_set;
5692                 len -= bits_to_set;
5693                 bits_to_set = BITS_PER_BYTE;
5694                 mask_to_set = ~0;
5695                 p++;
5696         }
5697         if (len) {
5698                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5699                 *p |= mask_to_set;
5700         }
5701 }
5702
5703 void le_bitmap_clear(u8 *map, unsigned int start, int len)
5704 {
5705         u8 *p = map + BIT_BYTE(start);
5706         const unsigned int size = start + len;
5707         int bits_to_clear = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5708         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(start);
5709
5710         while (len - bits_to_clear >= 0) {
5711                 *p &= ~mask_to_clear;
5712                 len -= bits_to_clear;
5713                 bits_to_clear = BITS_PER_BYTE;
5714                 mask_to_clear = ~0;
5715                 p++;
5716         }
5717         if (len) {
5718                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5719                 *p &= ~mask_to_clear;
5720         }
5721 }
5722
5723 /*
5724  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5725  * given bit number
5726  * @eb: the extent buffer
5727  * @start: offset of the bitmap item in the extent buffer
5728  * @nr: bit number
5729  * @page_index: return index of the page in the extent buffer that contains the
5730  * given bit number
5731  * @page_offset: return offset into the page given by page_index
5732  *
5733  * This helper hides the ugliness of finding the byte in an extent buffer which
5734  * contains a given bit.
5735  */
5736 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5737                                     unsigned long start, unsigned long nr,
5738                                     unsigned long *page_index,
5739                                     size_t *page_offset)
5740 {
5741         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5742         size_t byte_offset = BIT_BYTE(nr);
5743         size_t offset;
5744
5745         /*
5746          * The byte we want is the offset of the extent buffer + the offset of
5747          * the bitmap item in the extent buffer + the offset of the byte in the
5748          * bitmap item.
5749          */
5750         offset = start_offset + start + byte_offset;
5751
5752         *page_index = offset >> PAGE_SHIFT;
5753         *page_offset = offset & (PAGE_SIZE - 1);
5754 }
5755
5756 /**
5757  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5758  * @eb: the extent buffer
5759  * @start: offset of the bitmap item in the extent buffer
5760  * @nr: bit number to test
5761  */
5762 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5763                            unsigned long nr)
5764 {
5765         u8 *kaddr;
5766         struct page *page;
5767         unsigned long i;
5768         size_t offset;
5769
5770         eb_bitmap_offset(eb, start, nr, &i, &offset);
5771         page = eb->pages[i];
5772         WARN_ON(!PageUptodate(page));
5773         kaddr = page_address(page);
5774         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5775 }
5776
5777 /**
5778  * extent_buffer_bitmap_set - set an area of a bitmap
5779  * @eb: the extent buffer
5780  * @start: offset of the bitmap item in the extent buffer
5781  * @pos: bit number of the first bit
5782  * @len: number of bits to set
5783  */
5784 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5785                               unsigned long pos, unsigned long len)
5786 {
5787         u8 *kaddr;
5788         struct page *page;
5789         unsigned long i;
5790         size_t offset;
5791         const unsigned int size = pos + len;
5792         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5793         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5794
5795         eb_bitmap_offset(eb, start, pos, &i, &offset);
5796         page = eb->pages[i];
5797         WARN_ON(!PageUptodate(page));
5798         kaddr = page_address(page);
5799
5800         while (len >= bits_to_set) {
5801                 kaddr[offset] |= mask_to_set;
5802                 len -= bits_to_set;
5803                 bits_to_set = BITS_PER_BYTE;
5804                 mask_to_set = ~0;
5805                 if (++offset >= PAGE_SIZE && len > 0) {
5806                         offset = 0;
5807                         page = eb->pages[++i];
5808                         WARN_ON(!PageUptodate(page));
5809                         kaddr = page_address(page);
5810                 }
5811         }
5812         if (len) {
5813                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5814                 kaddr[offset] |= mask_to_set;
5815         }
5816 }
5817
5818
5819 /**
5820  * extent_buffer_bitmap_clear - clear an area of a bitmap
5821  * @eb: the extent buffer
5822  * @start: offset of the bitmap item in the extent buffer
5823  * @pos: bit number of the first bit
5824  * @len: number of bits to clear
5825  */
5826 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5827                                 unsigned long pos, unsigned long len)
5828 {
5829         u8 *kaddr;
5830         struct page *page;
5831         unsigned long i;
5832         size_t offset;
5833         const unsigned int size = pos + len;
5834         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5835         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5836
5837         eb_bitmap_offset(eb, start, pos, &i, &offset);
5838         page = eb->pages[i];
5839         WARN_ON(!PageUptodate(page));
5840         kaddr = page_address(page);
5841
5842         while (len >= bits_to_clear) {
5843                 kaddr[offset] &= ~mask_to_clear;
5844                 len -= bits_to_clear;
5845                 bits_to_clear = BITS_PER_BYTE;
5846                 mask_to_clear = ~0;
5847                 if (++offset >= PAGE_SIZE && len > 0) {
5848                         offset = 0;
5849                         page = eb->pages[++i];
5850                         WARN_ON(!PageUptodate(page));
5851                         kaddr = page_address(page);
5852                 }
5853         }
5854         if (len) {
5855                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5856                 kaddr[offset] &= ~mask_to_clear;
5857         }
5858 }
5859
5860 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5861 {
5862         unsigned long distance = (src > dst) ? src - dst : dst - src;
5863         return distance < len;
5864 }
5865
5866 static void copy_pages(struct page *dst_page, struct page *src_page,
5867                        unsigned long dst_off, unsigned long src_off,
5868                        unsigned long len)
5869 {
5870         char *dst_kaddr = page_address(dst_page);
5871         char *src_kaddr;
5872         int must_memmove = 0;
5873
5874         if (dst_page != src_page) {
5875                 src_kaddr = page_address(src_page);
5876         } else {
5877                 src_kaddr = dst_kaddr;
5878                 if (areas_overlap(src_off, dst_off, len))
5879                         must_memmove = 1;
5880         }
5881
5882         if (must_memmove)
5883                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5884         else
5885                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5886 }
5887
5888 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5889                            unsigned long src_offset, unsigned long len)
5890 {
5891         struct btrfs_fs_info *fs_info = dst->fs_info;
5892         size_t cur;
5893         size_t dst_off_in_page;
5894         size_t src_off_in_page;
5895         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5896         unsigned long dst_i;
5897         unsigned long src_i;
5898
5899         if (src_offset + len > dst->len) {
5900                 btrfs_err(fs_info,
5901                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5902                          src_offset, len, dst->len);
5903                 BUG_ON(1);
5904         }
5905         if (dst_offset + len > dst->len) {
5906                 btrfs_err(fs_info,
5907                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5908                          dst_offset, len, dst->len);
5909                 BUG_ON(1);
5910         }
5911
5912         while (len > 0) {
5913                 dst_off_in_page = (start_offset + dst_offset) &
5914                         (PAGE_SIZE - 1);
5915                 src_off_in_page = (start_offset + src_offset) &
5916                         (PAGE_SIZE - 1);
5917
5918                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5919                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5920
5921                 cur = min(len, (unsigned long)(PAGE_SIZE -
5922                                                src_off_in_page));
5923                 cur = min_t(unsigned long, cur,
5924                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5925
5926                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5927                            dst_off_in_page, src_off_in_page, cur);
5928
5929                 src_offset += cur;
5930                 dst_offset += cur;
5931                 len -= cur;
5932         }
5933 }
5934
5935 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5936                            unsigned long src_offset, unsigned long len)
5937 {
5938         struct btrfs_fs_info *fs_info = dst->fs_info;
5939         size_t cur;
5940         size_t dst_off_in_page;
5941         size_t src_off_in_page;
5942         unsigned long dst_end = dst_offset + len - 1;
5943         unsigned long src_end = src_offset + len - 1;
5944         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5945         unsigned long dst_i;
5946         unsigned long src_i;
5947
5948         if (src_offset + len > dst->len) {
5949                 btrfs_err(fs_info,
5950                           "memmove bogus src_offset %lu move len %lu len %lu",
5951                           src_offset, len, dst->len);
5952                 BUG_ON(1);
5953         }
5954         if (dst_offset + len > dst->len) {
5955                 btrfs_err(fs_info,
5956                           "memmove bogus dst_offset %lu move len %lu len %lu",
5957                           dst_offset, len, dst->len);
5958                 BUG_ON(1);
5959         }
5960         if (dst_offset < src_offset) {
5961                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5962                 return;
5963         }
5964         while (len > 0) {
5965                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5966                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5967
5968                 dst_off_in_page = (start_offset + dst_end) &
5969                         (PAGE_SIZE - 1);
5970                 src_off_in_page = (start_offset + src_end) &
5971                         (PAGE_SIZE - 1);
5972
5973                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5974                 cur = min(cur, dst_off_in_page + 1);
5975                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5976                            dst_off_in_page - cur + 1,
5977                            src_off_in_page - cur + 1, cur);
5978
5979                 dst_end -= cur;
5980                 src_end -= cur;
5981                 len -= cur;
5982         }
5983 }
5984
5985 int try_release_extent_buffer(struct page *page)
5986 {
5987         struct extent_buffer *eb;
5988
5989         /*
5990          * We need to make sure nobody is attaching this page to an eb right
5991          * now.
5992          */
5993         spin_lock(&page->mapping->private_lock);
5994         if (!PagePrivate(page)) {
5995                 spin_unlock(&page->mapping->private_lock);
5996                 return 1;
5997         }
5998
5999         eb = (struct extent_buffer *)page->private;
6000         BUG_ON(!eb);
6001
6002         /*
6003          * This is a little awful but should be ok, we need to make sure that
6004          * the eb doesn't disappear out from under us while we're looking at
6005          * this page.
6006          */
6007         spin_lock(&eb->refs_lock);
6008         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
6009                 spin_unlock(&eb->refs_lock);
6010                 spin_unlock(&page->mapping->private_lock);
6011                 return 0;
6012         }
6013         spin_unlock(&page->mapping->private_lock);
6014
6015         /*
6016          * If tree ref isn't set then we know the ref on this eb is a real ref,
6017          * so just return, this page will likely be freed soon anyway.
6018          */
6019         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
6020                 spin_unlock(&eb->refs_lock);
6021                 return 0;
6022         }
6023
6024         return release_extent_buffer(eb);
6025 }