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 bool 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 false;
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 false;
2319         }
2320
2321         return true;
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         if (!btrfs_check_repairable(inode, failed_bio, failrec,
2386                                     failed_mirror)) {
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         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2400
2401         btrfs_debug(btrfs_sb(inode->i_sb),
2402                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2403                 read_mode, failrec->this_mirror, failrec->in_validation);
2404
2405         status = tree->ops->submit_bio_hook(tree->private_data, bio, failrec->this_mirror,
2406                                          failrec->bio_flags, 0);
2407         if (status) {
2408                 free_io_failure(failure_tree, tree, failrec);
2409                 bio_put(bio);
2410                 ret = blk_status_to_errno(status);
2411         }
2412
2413         return ret;
2414 }
2415
2416 /* lots and lots of room for performance fixes in the end_bio funcs */
2417
2418 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2419 {
2420         int uptodate = (err == 0);
2421         struct extent_io_tree *tree;
2422         int ret = 0;
2423
2424         tree = &BTRFS_I(page->mapping->host)->io_tree;
2425
2426         if (tree->ops && tree->ops->writepage_end_io_hook)
2427                 tree->ops->writepage_end_io_hook(page, start, end, NULL,
2428                                 uptodate);
2429
2430         if (!uptodate) {
2431                 ClearPageUptodate(page);
2432                 SetPageError(page);
2433                 ret = err < 0 ? err : -EIO;
2434                 mapping_set_error(page->mapping, ret);
2435         }
2436 }
2437
2438 /*
2439  * after a writepage IO is done, we need to:
2440  * clear the uptodate bits on error
2441  * clear the writeback bits in the extent tree for this IO
2442  * end_page_writeback if the page has no more pending IO
2443  *
2444  * Scheduling is not allowed, so the extent state tree is expected
2445  * to have one and only one object corresponding to this IO.
2446  */
2447 static void end_bio_extent_writepage(struct bio *bio)
2448 {
2449         int error = blk_status_to_errno(bio->bi_status);
2450         struct bio_vec *bvec;
2451         u64 start;
2452         u64 end;
2453         int i;
2454
2455         ASSERT(!bio_flagged(bio, BIO_CLONED));
2456         bio_for_each_segment_all(bvec, bio, i) {
2457                 struct page *page = bvec->bv_page;
2458                 struct inode *inode = page->mapping->host;
2459                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2460
2461                 /* We always issue full-page reads, but if some block
2462                  * in a page fails to read, blk_update_request() will
2463                  * advance bv_offset and adjust bv_len to compensate.
2464                  * Print a warning for nonzero offsets, and an error
2465                  * if they don't add up to a full page.  */
2466                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2467                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2468                                 btrfs_err(fs_info,
2469                                    "partial page write in btrfs with offset %u and length %u",
2470                                         bvec->bv_offset, bvec->bv_len);
2471                         else
2472                                 btrfs_info(fs_info,
2473                                    "incomplete page write in btrfs with offset %u and length %u",
2474                                         bvec->bv_offset, bvec->bv_len);
2475                 }
2476
2477                 start = page_offset(page);
2478                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2479
2480                 end_extent_writepage(page, error, start, end);
2481                 end_page_writeback(page);
2482         }
2483
2484         bio_put(bio);
2485 }
2486
2487 static void
2488 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2489                               int uptodate)
2490 {
2491         struct extent_state *cached = NULL;
2492         u64 end = start + len - 1;
2493
2494         if (uptodate && tree->track_uptodate)
2495                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2496         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2497 }
2498
2499 /*
2500  * after a readpage IO is done, we need to:
2501  * clear the uptodate bits on error
2502  * set the uptodate bits if things worked
2503  * set the page up to date if all extents in the tree are uptodate
2504  * clear the lock bit in the extent tree
2505  * unlock the page if there are no other extents locked for it
2506  *
2507  * Scheduling is not allowed, so the extent state tree is expected
2508  * to have one and only one object corresponding to this IO.
2509  */
2510 static void end_bio_extent_readpage(struct bio *bio)
2511 {
2512         struct bio_vec *bvec;
2513         int uptodate = !bio->bi_status;
2514         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2515         struct extent_io_tree *tree, *failure_tree;
2516         u64 offset = 0;
2517         u64 start;
2518         u64 end;
2519         u64 len;
2520         u64 extent_start = 0;
2521         u64 extent_len = 0;
2522         int mirror;
2523         int ret;
2524         int i;
2525
2526         ASSERT(!bio_flagged(bio, BIO_CLONED));
2527         bio_for_each_segment_all(bvec, bio, i) {
2528                 struct page *page = bvec->bv_page;
2529                 struct inode *inode = page->mapping->host;
2530                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2531
2532                 btrfs_debug(fs_info,
2533                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2534                         (u64)bio->bi_iter.bi_sector, bio->bi_status,
2535                         io_bio->mirror_num);
2536                 tree = &BTRFS_I(inode)->io_tree;
2537                 failure_tree = &BTRFS_I(inode)->io_failure_tree;
2538
2539                 /* We always issue full-page reads, but if some block
2540                  * in a page fails to read, blk_update_request() will
2541                  * advance bv_offset and adjust bv_len to compensate.
2542                  * Print a warning for nonzero offsets, and an error
2543                  * if they don't add up to a full page.  */
2544                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2545                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2546                                 btrfs_err(fs_info,
2547                                         "partial page read in btrfs with offset %u and length %u",
2548                                         bvec->bv_offset, bvec->bv_len);
2549                         else
2550                                 btrfs_info(fs_info,
2551                                         "incomplete page read in btrfs with offset %u and length %u",
2552                                         bvec->bv_offset, bvec->bv_len);
2553                 }
2554
2555                 start = page_offset(page);
2556                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2557                 len = bvec->bv_len;
2558
2559                 mirror = io_bio->mirror_num;
2560                 if (likely(uptodate && tree->ops)) {
2561                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2562                                                               page, start, end,
2563                                                               mirror);
2564                         if (ret)
2565                                 uptodate = 0;
2566                         else
2567                                 clean_io_failure(BTRFS_I(inode)->root->fs_info,
2568                                                  failure_tree, tree, start,
2569                                                  page,
2570                                                  btrfs_ino(BTRFS_I(inode)), 0);
2571                 }
2572
2573                 if (likely(uptodate))
2574                         goto readpage_ok;
2575
2576                 if (tree->ops) {
2577                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2578                         if (ret == -EAGAIN) {
2579                                 /*
2580                                  * Data inode's readpage_io_failed_hook() always
2581                                  * returns -EAGAIN.
2582                                  *
2583                                  * The generic bio_readpage_error handles errors
2584                                  * the following way: If possible, new read
2585                                  * requests are created and submitted and will
2586                                  * end up in end_bio_extent_readpage as well (if
2587                                  * we're lucky, not in the !uptodate case). In
2588                                  * that case it returns 0 and we just go on with
2589                                  * the next page in our bio. If it can't handle
2590                                  * the error it will return -EIO and we remain
2591                                  * responsible for that page.
2592                                  */
2593                                 ret = bio_readpage_error(bio, offset, page,
2594                                                          start, end, mirror);
2595                                 if (ret == 0) {
2596                                         uptodate = !bio->bi_status;
2597                                         offset += len;
2598                                         continue;
2599                                 }
2600                         }
2601
2602                         /*
2603                          * metadata's readpage_io_failed_hook() always returns
2604                          * -EIO and fixes nothing.  -EIO is also returned if
2605                          * data inode error could not be fixed.
2606                          */
2607                         ASSERT(ret == -EIO);
2608                 }
2609 readpage_ok:
2610                 if (likely(uptodate)) {
2611                         loff_t i_size = i_size_read(inode);
2612                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2613                         unsigned off;
2614
2615                         /* Zero out the end if this page straddles i_size */
2616                         off = i_size & (PAGE_SIZE-1);
2617                         if (page->index == end_index && off)
2618                                 zero_user_segment(page, off, PAGE_SIZE);
2619                         SetPageUptodate(page);
2620                 } else {
2621                         ClearPageUptodate(page);
2622                         SetPageError(page);
2623                 }
2624                 unlock_page(page);
2625                 offset += len;
2626
2627                 if (unlikely(!uptodate)) {
2628                         if (extent_len) {
2629                                 endio_readpage_release_extent(tree,
2630                                                               extent_start,
2631                                                               extent_len, 1);
2632                                 extent_start = 0;
2633                                 extent_len = 0;
2634                         }
2635                         endio_readpage_release_extent(tree, start,
2636                                                       end - start + 1, 0);
2637                 } else if (!extent_len) {
2638                         extent_start = start;
2639                         extent_len = end + 1 - start;
2640                 } else if (extent_start + extent_len == start) {
2641                         extent_len += end + 1 - start;
2642                 } else {
2643                         endio_readpage_release_extent(tree, extent_start,
2644                                                       extent_len, uptodate);
2645                         extent_start = start;
2646                         extent_len = end + 1 - start;
2647                 }
2648         }
2649
2650         if (extent_len)
2651                 endio_readpage_release_extent(tree, extent_start, extent_len,
2652                                               uptodate);
2653         if (io_bio->end_io)
2654                 io_bio->end_io(io_bio, blk_status_to_errno(bio->bi_status));
2655         bio_put(bio);
2656 }
2657
2658 /*
2659  * Initialize the members up to but not including 'bio'. Use after allocating a
2660  * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
2661  * 'bio' because use of __GFP_ZERO is not supported.
2662  */
2663 static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
2664 {
2665         memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
2666 }
2667
2668 /*
2669  * The following helpers allocate a bio. As it's backed by a bioset, it'll
2670  * never fail.  We're returning a bio right now but you can call btrfs_io_bio
2671  * for the appropriate container_of magic
2672  */
2673 struct bio *btrfs_bio_alloc(struct block_device *bdev, u64 first_byte)
2674 {
2675         struct bio *bio;
2676
2677         bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, btrfs_bioset);
2678         bio->bi_bdev = bdev;
2679         bio->bi_iter.bi_sector = first_byte >> 9;
2680         btrfs_io_bio_init(btrfs_io_bio(bio));
2681         return bio;
2682 }
2683
2684 struct bio *btrfs_bio_clone(struct bio *bio)
2685 {
2686         struct btrfs_io_bio *btrfs_bio;
2687         struct bio *new;
2688
2689         /* Bio allocation backed by a bioset does not fail */
2690         new = bio_clone_fast(bio, GFP_NOFS, btrfs_bioset);
2691         btrfs_bio = btrfs_io_bio(new);
2692         btrfs_io_bio_init(btrfs_bio);
2693         btrfs_bio->iter = bio->bi_iter;
2694         return new;
2695 }
2696
2697 struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
2698 {
2699         struct bio *bio;
2700
2701         /* Bio allocation backed by a bioset does not fail */
2702         bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, btrfs_bioset);
2703         btrfs_io_bio_init(btrfs_io_bio(bio));
2704         return bio;
2705 }
2706
2707 struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
2708 {
2709         struct bio *bio;
2710         struct btrfs_io_bio *btrfs_bio;
2711
2712         /* this will never fail when it's backed by a bioset */
2713         bio = bio_clone_fast(orig, GFP_NOFS, btrfs_bioset);
2714         ASSERT(bio);
2715
2716         btrfs_bio = btrfs_io_bio(bio);
2717         btrfs_io_bio_init(btrfs_bio);
2718
2719         bio_trim(bio, offset >> 9, size >> 9);
2720         btrfs_bio->iter = bio->bi_iter;
2721         return bio;
2722 }
2723
2724 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2725                                        unsigned long bio_flags)
2726 {
2727         blk_status_t ret = 0;
2728         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2729         struct page *page = bvec->bv_page;
2730         struct extent_io_tree *tree = bio->bi_private;
2731         u64 start;
2732
2733         start = page_offset(page) + bvec->bv_offset;
2734
2735         bio->bi_private = NULL;
2736         bio_get(bio);
2737
2738         if (tree->ops)
2739                 ret = tree->ops->submit_bio_hook(tree->private_data, bio,
2740                                            mirror_num, bio_flags, start);
2741         else
2742                 btrfsic_submit_bio(bio);
2743
2744         bio_put(bio);
2745         return blk_status_to_errno(ret);
2746 }
2747
2748 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2749                      unsigned long offset, size_t size, struct bio *bio,
2750                      unsigned long bio_flags)
2751 {
2752         int ret = 0;
2753         if (tree->ops)
2754                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2755                                                 bio_flags);
2756         return ret;
2757
2758 }
2759
2760 static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree,
2761                               struct writeback_control *wbc,
2762                               struct page *page, sector_t sector,
2763                               size_t size, unsigned long offset,
2764                               struct block_device *bdev,
2765                               struct bio **bio_ret,
2766                               bio_end_io_t end_io_func,
2767                               int mirror_num,
2768                               unsigned long prev_bio_flags,
2769                               unsigned long bio_flags,
2770                               bool force_bio_submit)
2771 {
2772         int ret = 0;
2773         struct bio *bio;
2774         int contig = 0;
2775         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2776         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2777
2778         if (bio_ret && *bio_ret) {
2779                 bio = *bio_ret;
2780                 if (old_compressed)
2781                         contig = bio->bi_iter.bi_sector == sector;
2782                 else
2783                         contig = bio_end_sector(bio) == sector;
2784
2785                 if (prev_bio_flags != bio_flags || !contig ||
2786                     force_bio_submit ||
2787                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2788                     bio_add_page(bio, page, page_size, offset) < page_size) {
2789                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2790                         if (ret < 0) {
2791                                 *bio_ret = NULL;
2792                                 return ret;
2793                         }
2794                         bio = NULL;
2795                 } else {
2796                         if (wbc)
2797                                 wbc_account_io(wbc, page, page_size);
2798                         return 0;
2799                 }
2800         }
2801
2802         bio = btrfs_bio_alloc(bdev, sector << 9);
2803         bio_add_page(bio, page, page_size, offset);
2804         bio->bi_end_io = end_io_func;
2805         bio->bi_private = tree;
2806         bio->bi_write_hint = page->mapping->host->i_write_hint;
2807         bio_set_op_attrs(bio, op, op_flags);
2808         if (wbc) {
2809                 wbc_init_bio(wbc, bio);
2810                 wbc_account_io(wbc, page, page_size);
2811         }
2812
2813         if (bio_ret)
2814                 *bio_ret = bio;
2815         else
2816                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2817
2818         return ret;
2819 }
2820
2821 static void attach_extent_buffer_page(struct extent_buffer *eb,
2822                                       struct page *page)
2823 {
2824         if (!PagePrivate(page)) {
2825                 SetPagePrivate(page);
2826                 get_page(page);
2827                 set_page_private(page, (unsigned long)eb);
2828         } else {
2829                 WARN_ON(page->private != (unsigned long)eb);
2830         }
2831 }
2832
2833 void set_page_extent_mapped(struct page *page)
2834 {
2835         if (!PagePrivate(page)) {
2836                 SetPagePrivate(page);
2837                 get_page(page);
2838                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2839         }
2840 }
2841
2842 static struct extent_map *
2843 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2844                  u64 start, u64 len, get_extent_t *get_extent,
2845                  struct extent_map **em_cached)
2846 {
2847         struct extent_map *em;
2848
2849         if (em_cached && *em_cached) {
2850                 em = *em_cached;
2851                 if (extent_map_in_tree(em) && start >= em->start &&
2852                     start < extent_map_end(em)) {
2853                         refcount_inc(&em->refs);
2854                         return em;
2855                 }
2856
2857                 free_extent_map(em);
2858                 *em_cached = NULL;
2859         }
2860
2861         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2862         if (em_cached && !IS_ERR_OR_NULL(em)) {
2863                 BUG_ON(*em_cached);
2864                 refcount_inc(&em->refs);
2865                 *em_cached = em;
2866         }
2867         return em;
2868 }
2869 /*
2870  * basic readpage implementation.  Locked extent state structs are inserted
2871  * into the tree that are removed when the IO is done (by the end_io
2872  * handlers)
2873  * XXX JDM: This needs looking at to ensure proper page locking
2874  * return 0 on success, otherwise return error
2875  */
2876 static int __do_readpage(struct extent_io_tree *tree,
2877                          struct page *page,
2878                          get_extent_t *get_extent,
2879                          struct extent_map **em_cached,
2880                          struct bio **bio, int mirror_num,
2881                          unsigned long *bio_flags, int read_flags,
2882                          u64 *prev_em_start)
2883 {
2884         struct inode *inode = page->mapping->host;
2885         u64 start = page_offset(page);
2886         u64 page_end = start + PAGE_SIZE - 1;
2887         u64 end;
2888         u64 cur = start;
2889         u64 extent_offset;
2890         u64 last_byte = i_size_read(inode);
2891         u64 block_start;
2892         u64 cur_end;
2893         sector_t sector;
2894         struct extent_map *em;
2895         struct block_device *bdev;
2896         int ret = 0;
2897         int nr = 0;
2898         size_t pg_offset = 0;
2899         size_t iosize;
2900         size_t disk_io_size;
2901         size_t blocksize = inode->i_sb->s_blocksize;
2902         unsigned long this_bio_flag = 0;
2903
2904         set_page_extent_mapped(page);
2905
2906         end = page_end;
2907         if (!PageUptodate(page)) {
2908                 if (cleancache_get_page(page) == 0) {
2909                         BUG_ON(blocksize != PAGE_SIZE);
2910                         unlock_extent(tree, start, end);
2911                         goto out;
2912                 }
2913         }
2914
2915         if (page->index == last_byte >> PAGE_SHIFT) {
2916                 char *userpage;
2917                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2918
2919                 if (zero_offset) {
2920                         iosize = PAGE_SIZE - zero_offset;
2921                         userpage = kmap_atomic(page);
2922                         memset(userpage + zero_offset, 0, iosize);
2923                         flush_dcache_page(page);
2924                         kunmap_atomic(userpage);
2925                 }
2926         }
2927         while (cur <= end) {
2928                 bool force_bio_submit = false;
2929
2930                 if (cur >= last_byte) {
2931                         char *userpage;
2932                         struct extent_state *cached = NULL;
2933
2934                         iosize = PAGE_SIZE - pg_offset;
2935                         userpage = kmap_atomic(page);
2936                         memset(userpage + pg_offset, 0, iosize);
2937                         flush_dcache_page(page);
2938                         kunmap_atomic(userpage);
2939                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2940                                             &cached, GFP_NOFS);
2941                         unlock_extent_cached(tree, cur,
2942                                              cur + iosize - 1,
2943                                              &cached, GFP_NOFS);
2944                         break;
2945                 }
2946                 em = __get_extent_map(inode, page, pg_offset, cur,
2947                                       end - cur + 1, get_extent, em_cached);
2948                 if (IS_ERR_OR_NULL(em)) {
2949                         SetPageError(page);
2950                         unlock_extent(tree, cur, end);
2951                         break;
2952                 }
2953                 extent_offset = cur - em->start;
2954                 BUG_ON(extent_map_end(em) <= cur);
2955                 BUG_ON(end < cur);
2956
2957                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2958                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2959                         extent_set_compress_type(&this_bio_flag,
2960                                                  em->compress_type);
2961                 }
2962
2963                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2964                 cur_end = min(extent_map_end(em) - 1, end);
2965                 iosize = ALIGN(iosize, blocksize);
2966                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2967                         disk_io_size = em->block_len;
2968                         sector = em->block_start >> 9;
2969                 } else {
2970                         sector = (em->block_start + extent_offset) >> 9;
2971                         disk_io_size = iosize;
2972                 }
2973                 bdev = em->bdev;
2974                 block_start = em->block_start;
2975                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2976                         block_start = EXTENT_MAP_HOLE;
2977
2978                 /*
2979                  * If we have a file range that points to a compressed extent
2980                  * and it's followed by a consecutive file range that points to
2981                  * to the same compressed extent (possibly with a different
2982                  * offset and/or length, so it either points to the whole extent
2983                  * or only part of it), we must make sure we do not submit a
2984                  * single bio to populate the pages for the 2 ranges because
2985                  * this makes the compressed extent read zero out the pages
2986                  * belonging to the 2nd range. Imagine the following scenario:
2987                  *
2988                  *  File layout
2989                  *  [0 - 8K]                     [8K - 24K]
2990                  *    |                               |
2991                  *    |                               |
2992                  * points to extent X,         points to extent X,
2993                  * offset 4K, length of 8K     offset 0, length 16K
2994                  *
2995                  * [extent X, compressed length = 4K uncompressed length = 16K]
2996                  *
2997                  * If the bio to read the compressed extent covers both ranges,
2998                  * it will decompress extent X into the pages belonging to the
2999                  * first range and then it will stop, zeroing out the remaining
3000                  * pages that belong to the other range that points to extent X.
3001                  * So here we make sure we submit 2 bios, one for the first
3002                  * range and another one for the third range. Both will target
3003                  * the same physical extent from disk, but we can't currently
3004                  * make the compressed bio endio callback populate the pages
3005                  * for both ranges because each compressed bio is tightly
3006                  * coupled with a single extent map, and each range can have
3007                  * an extent map with a different offset value relative to the
3008                  * uncompressed data of our extent and different lengths. This
3009                  * is a corner case so we prioritize correctness over
3010                  * non-optimal behavior (submitting 2 bios for the same extent).
3011                  */
3012                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3013                     prev_em_start && *prev_em_start != (u64)-1 &&
3014                     *prev_em_start != em->orig_start)
3015                         force_bio_submit = true;
3016
3017                 if (prev_em_start)
3018                         *prev_em_start = em->orig_start;
3019
3020                 free_extent_map(em);
3021                 em = NULL;
3022
3023                 /* we've found a hole, just zero and go on */
3024                 if (block_start == EXTENT_MAP_HOLE) {
3025                         char *userpage;
3026                         struct extent_state *cached = NULL;
3027
3028                         userpage = kmap_atomic(page);
3029                         memset(userpage + pg_offset, 0, iosize);
3030                         flush_dcache_page(page);
3031                         kunmap_atomic(userpage);
3032
3033                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3034                                             &cached, GFP_NOFS);
3035                         unlock_extent_cached(tree, cur,
3036                                              cur + iosize - 1,
3037                                              &cached, GFP_NOFS);
3038                         cur = cur + iosize;
3039                         pg_offset += iosize;
3040                         continue;
3041                 }
3042                 /* the get_extent function already copied into the page */
3043                 if (test_range_bit(tree, cur, cur_end,
3044                                    EXTENT_UPTODATE, 1, NULL)) {
3045                         check_page_uptodate(tree, page);
3046                         unlock_extent(tree, cur, cur + iosize - 1);
3047                         cur = cur + iosize;
3048                         pg_offset += iosize;
3049                         continue;
3050                 }
3051                 /* we have an inline extent but it didn't get marked up
3052                  * to date.  Error out
3053                  */
3054                 if (block_start == EXTENT_MAP_INLINE) {
3055                         SetPageError(page);
3056                         unlock_extent(tree, cur, cur + iosize - 1);
3057                         cur = cur + iosize;
3058                         pg_offset += iosize;
3059                         continue;
3060                 }
3061
3062                 ret = submit_extent_page(REQ_OP_READ, read_flags, tree, NULL,
3063                                          page, sector, disk_io_size, pg_offset,
3064                                          bdev, bio,
3065                                          end_bio_extent_readpage, mirror_num,
3066                                          *bio_flags,
3067                                          this_bio_flag,
3068                                          force_bio_submit);
3069                 if (!ret) {
3070                         nr++;
3071                         *bio_flags = this_bio_flag;
3072                 } else {
3073                         SetPageError(page);
3074                         unlock_extent(tree, cur, cur + iosize - 1);
3075                         goto out;
3076                 }
3077                 cur = cur + iosize;
3078                 pg_offset += iosize;
3079         }
3080 out:
3081         if (!nr) {
3082                 if (!PageError(page))
3083                         SetPageUptodate(page);
3084                 unlock_page(page);
3085         }
3086         return ret;
3087 }
3088
3089 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3090                                              struct page *pages[], int nr_pages,
3091                                              u64 start, u64 end,
3092                                              get_extent_t *get_extent,
3093                                              struct extent_map **em_cached,
3094                                              struct bio **bio, int mirror_num,
3095                                              unsigned long *bio_flags,
3096                                              u64 *prev_em_start)
3097 {
3098         struct inode *inode;
3099         struct btrfs_ordered_extent *ordered;
3100         int index;
3101
3102         inode = pages[0]->mapping->host;
3103         while (1) {
3104                 lock_extent(tree, start, end);
3105                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3106                                                      end - start + 1);
3107                 if (!ordered)
3108                         break;
3109                 unlock_extent(tree, start, end);
3110                 btrfs_start_ordered_extent(inode, ordered, 1);
3111                 btrfs_put_ordered_extent(ordered);
3112         }
3113
3114         for (index = 0; index < nr_pages; index++) {
3115                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3116                               mirror_num, bio_flags, 0, prev_em_start);
3117                 put_page(pages[index]);
3118         }
3119 }
3120
3121 static void __extent_readpages(struct extent_io_tree *tree,
3122                                struct page *pages[],
3123                                int nr_pages, get_extent_t *get_extent,
3124                                struct extent_map **em_cached,
3125                                struct bio **bio, int mirror_num,
3126                                unsigned long *bio_flags,
3127                                u64 *prev_em_start)
3128 {
3129         u64 start = 0;
3130         u64 end = 0;
3131         u64 page_start;
3132         int index;
3133         int first_index = 0;
3134
3135         for (index = 0; index < nr_pages; index++) {
3136                 page_start = page_offset(pages[index]);
3137                 if (!end) {
3138                         start = page_start;
3139                         end = start + PAGE_SIZE - 1;
3140                         first_index = index;
3141                 } else if (end + 1 == page_start) {
3142                         end += PAGE_SIZE;
3143                 } else {
3144                         __do_contiguous_readpages(tree, &pages[first_index],
3145                                                   index - first_index, start,
3146                                                   end, get_extent, em_cached,
3147                                                   bio, mirror_num, bio_flags,
3148                                                   prev_em_start);
3149                         start = page_start;
3150                         end = start + PAGE_SIZE - 1;
3151                         first_index = index;
3152                 }
3153         }
3154
3155         if (end)
3156                 __do_contiguous_readpages(tree, &pages[first_index],
3157                                           index - first_index, start,
3158                                           end, get_extent, em_cached, bio,
3159                                           mirror_num, bio_flags,
3160                                           prev_em_start);
3161 }
3162
3163 static int __extent_read_full_page(struct extent_io_tree *tree,
3164                                    struct page *page,
3165                                    get_extent_t *get_extent,
3166                                    struct bio **bio, int mirror_num,
3167                                    unsigned long *bio_flags, int read_flags)
3168 {
3169         struct inode *inode = page->mapping->host;
3170         struct btrfs_ordered_extent *ordered;
3171         u64 start = page_offset(page);
3172         u64 end = start + PAGE_SIZE - 1;
3173         int ret;
3174
3175         while (1) {
3176                 lock_extent(tree, start, end);
3177                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3178                                                 PAGE_SIZE);
3179                 if (!ordered)
3180                         break;
3181                 unlock_extent(tree, start, end);
3182                 btrfs_start_ordered_extent(inode, ordered, 1);
3183                 btrfs_put_ordered_extent(ordered);
3184         }
3185
3186         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3187                             bio_flags, read_flags, NULL);
3188         return ret;
3189 }
3190
3191 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3192                             get_extent_t *get_extent, int mirror_num)
3193 {
3194         struct bio *bio = NULL;
3195         unsigned long bio_flags = 0;
3196         int ret;
3197
3198         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3199                                       &bio_flags, 0);
3200         if (bio)
3201                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3202         return ret;
3203 }
3204
3205 static void update_nr_written(struct writeback_control *wbc,
3206                               unsigned long nr_written)
3207 {
3208         wbc->nr_to_write -= nr_written;
3209 }
3210
3211 /*
3212  * helper for __extent_writepage, doing all of the delayed allocation setup.
3213  *
3214  * This returns 1 if our fill_delalloc function did all the work required
3215  * to write the page (copy into inline extent).  In this case the IO has
3216  * been started and the page is already unlocked.
3217  *
3218  * This returns 0 if all went well (page still locked)
3219  * This returns < 0 if there were errors (page still locked)
3220  */
3221 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3222                               struct page *page, struct writeback_control *wbc,
3223                               struct extent_page_data *epd,
3224                               u64 delalloc_start,
3225                               unsigned long *nr_written)
3226 {
3227         struct extent_io_tree *tree = epd->tree;
3228         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3229         u64 nr_delalloc;
3230         u64 delalloc_to_write = 0;
3231         u64 delalloc_end = 0;
3232         int ret;
3233         int page_started = 0;
3234
3235         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3236                 return 0;
3237
3238         while (delalloc_end < page_end) {
3239                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3240                                                page,
3241                                                &delalloc_start,
3242                                                &delalloc_end,
3243                                                BTRFS_MAX_EXTENT_SIZE);
3244                 if (nr_delalloc == 0) {
3245                         delalloc_start = delalloc_end + 1;
3246                         continue;
3247                 }
3248                 ret = tree->ops->fill_delalloc(inode, page,
3249                                                delalloc_start,
3250                                                delalloc_end,
3251                                                &page_started,
3252                                                nr_written);
3253                 /* File system has been set read-only */
3254                 if (ret) {
3255                         SetPageError(page);
3256                         /* fill_delalloc should be return < 0 for error
3257                          * but just in case, we use > 0 here meaning the
3258                          * IO is started, so we don't want to return > 0
3259                          * unless things are going well.
3260                          */
3261                         ret = ret < 0 ? ret : -EIO;
3262                         goto done;
3263                 }
3264                 /*
3265                  * delalloc_end is already one less than the total length, so
3266                  * we don't subtract one from PAGE_SIZE
3267                  */
3268                 delalloc_to_write += (delalloc_end - delalloc_start +
3269                                       PAGE_SIZE) >> PAGE_SHIFT;
3270                 delalloc_start = delalloc_end + 1;
3271         }
3272         if (wbc->nr_to_write < delalloc_to_write) {
3273                 int thresh = 8192;
3274
3275                 if (delalloc_to_write < thresh * 2)
3276                         thresh = delalloc_to_write;
3277                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3278                                          thresh);
3279         }
3280
3281         /* did the fill delalloc function already unlock and start
3282          * the IO?
3283          */
3284         if (page_started) {
3285                 /*
3286                  * we've unlocked the page, so we can't update
3287                  * the mapping's writeback index, just update
3288                  * nr_to_write.
3289                  */
3290                 wbc->nr_to_write -= *nr_written;
3291                 return 1;
3292         }
3293
3294         ret = 0;
3295
3296 done:
3297         return ret;
3298 }
3299
3300 /*
3301  * helper for __extent_writepage.  This calls the writepage start hooks,
3302  * and does the loop to map the page into extents and bios.
3303  *
3304  * We return 1 if the IO is started and the page is unlocked,
3305  * 0 if all went well (page still locked)
3306  * < 0 if there were errors (page still locked)
3307  */
3308 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3309                                  struct page *page,
3310                                  struct writeback_control *wbc,
3311                                  struct extent_page_data *epd,
3312                                  loff_t i_size,
3313                                  unsigned long nr_written,
3314                                  int write_flags, int *nr_ret)
3315 {
3316         struct extent_io_tree *tree = epd->tree;
3317         u64 start = page_offset(page);
3318         u64 page_end = start + PAGE_SIZE - 1;
3319         u64 end;
3320         u64 cur = start;
3321         u64 extent_offset;
3322         u64 block_start;
3323         u64 iosize;
3324         sector_t sector;
3325         struct extent_map *em;
3326         struct block_device *bdev;
3327         size_t pg_offset = 0;
3328         size_t blocksize;
3329         int ret = 0;
3330         int nr = 0;
3331         bool compressed;
3332
3333         if (tree->ops && tree->ops->writepage_start_hook) {
3334                 ret = tree->ops->writepage_start_hook(page, start,
3335                                                       page_end);
3336                 if (ret) {
3337                         /* Fixup worker will requeue */
3338                         if (ret == -EBUSY)
3339                                 wbc->pages_skipped++;
3340                         else
3341                                 redirty_page_for_writepage(wbc, page);
3342
3343                         update_nr_written(wbc, nr_written);
3344                         unlock_page(page);
3345                         return 1;
3346                 }
3347         }
3348
3349         /*
3350          * we don't want to touch the inode after unlocking the page,
3351          * so we update the mapping writeback index now
3352          */
3353         update_nr_written(wbc, nr_written + 1);
3354
3355         end = page_end;
3356         if (i_size <= start) {
3357                 if (tree->ops && tree->ops->writepage_end_io_hook)
3358                         tree->ops->writepage_end_io_hook(page, start,
3359                                                          page_end, NULL, 1);
3360                 goto done;
3361         }
3362
3363         blocksize = inode->i_sb->s_blocksize;
3364
3365         while (cur <= end) {
3366                 u64 em_end;
3367
3368                 if (cur >= i_size) {
3369                         if (tree->ops && tree->ops->writepage_end_io_hook)
3370                                 tree->ops->writepage_end_io_hook(page, cur,
3371                                                          page_end, NULL, 1);
3372                         break;
3373                 }
3374                 em = epd->get_extent(BTRFS_I(inode), page, pg_offset, cur,
3375                                      end - cur + 1, 1);
3376                 if (IS_ERR_OR_NULL(em)) {
3377                         SetPageError(page);
3378                         ret = PTR_ERR_OR_ZERO(em);
3379                         break;
3380                 }
3381
3382                 extent_offset = cur - em->start;
3383                 em_end = extent_map_end(em);
3384                 BUG_ON(em_end <= cur);
3385                 BUG_ON(end < cur);
3386                 iosize = min(em_end - cur, end - cur + 1);
3387                 iosize = ALIGN(iosize, blocksize);
3388                 sector = (em->block_start + extent_offset) >> 9;
3389                 bdev = em->bdev;
3390                 block_start = em->block_start;
3391                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3392                 free_extent_map(em);
3393                 em = NULL;
3394
3395                 /*
3396                  * compressed and inline extents are written through other
3397                  * paths in the FS
3398                  */
3399                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3400                     block_start == EXTENT_MAP_INLINE) {
3401                         /*
3402                          * end_io notification does not happen here for
3403                          * compressed extents
3404                          */
3405                         if (!compressed && tree->ops &&
3406                             tree->ops->writepage_end_io_hook)
3407                                 tree->ops->writepage_end_io_hook(page, cur,
3408                                                          cur + iosize - 1,
3409                                                          NULL, 1);
3410                         else if (compressed) {
3411                                 /* we don't want to end_page_writeback on
3412                                  * a compressed extent.  this happens
3413                                  * elsewhere
3414                                  */
3415                                 nr++;
3416                         }
3417
3418                         cur += iosize;
3419                         pg_offset += iosize;
3420                         continue;
3421                 }
3422
3423                 set_range_writeback(tree, cur, cur + iosize - 1);
3424                 if (!PageWriteback(page)) {
3425                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3426                                    "page %lu not writeback, cur %llu end %llu",
3427                                page->index, cur, end);
3428                 }
3429
3430                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3431                                          page, sector, iosize, pg_offset,
3432                                          bdev, &epd->bio,
3433                                          end_bio_extent_writepage,
3434                                          0, 0, 0, false);
3435                 if (ret) {
3436                         SetPageError(page);
3437                         if (PageWriteback(page))
3438                                 end_page_writeback(page);
3439                 }
3440
3441                 cur = cur + iosize;
3442                 pg_offset += iosize;
3443                 nr++;
3444         }
3445 done:
3446         *nr_ret = nr;
3447         return ret;
3448 }
3449
3450 /*
3451  * the writepage semantics are similar to regular writepage.  extent
3452  * records are inserted to lock ranges in the tree, and as dirty areas
3453  * are found, they are marked writeback.  Then the lock bits are removed
3454  * and the end_io handler clears the writeback ranges
3455  */
3456 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3457                               void *data)
3458 {
3459         struct inode *inode = page->mapping->host;
3460         struct extent_page_data *epd = data;
3461         u64 start = page_offset(page);
3462         u64 page_end = start + PAGE_SIZE - 1;
3463         int ret;
3464         int nr = 0;
3465         size_t pg_offset = 0;
3466         loff_t i_size = i_size_read(inode);
3467         unsigned long end_index = i_size >> PAGE_SHIFT;
3468         int write_flags = 0;
3469         unsigned long nr_written = 0;
3470
3471         if (wbc->sync_mode == WB_SYNC_ALL)
3472                 write_flags = REQ_SYNC;
3473
3474         trace___extent_writepage(page, inode, wbc);
3475
3476         WARN_ON(!PageLocked(page));
3477
3478         ClearPageError(page);
3479
3480         pg_offset = i_size & (PAGE_SIZE - 1);
3481         if (page->index > end_index ||
3482            (page->index == end_index && !pg_offset)) {
3483                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3484                 unlock_page(page);
3485                 return 0;
3486         }
3487
3488         if (page->index == end_index) {
3489                 char *userpage;
3490
3491                 userpage = kmap_atomic(page);
3492                 memset(userpage + pg_offset, 0,
3493                        PAGE_SIZE - pg_offset);
3494                 kunmap_atomic(userpage);
3495                 flush_dcache_page(page);
3496         }
3497
3498         pg_offset = 0;
3499
3500         set_page_extent_mapped(page);
3501
3502         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3503         if (ret == 1)
3504                 goto done_unlocked;
3505         if (ret)
3506                 goto done;
3507
3508         ret = __extent_writepage_io(inode, page, wbc, epd,
3509                                     i_size, nr_written, write_flags, &nr);
3510         if (ret == 1)
3511                 goto done_unlocked;
3512
3513 done:
3514         if (nr == 0) {
3515                 /* make sure the mapping tag for page dirty gets cleared */
3516                 set_page_writeback(page);
3517                 end_page_writeback(page);
3518         }
3519         if (PageError(page)) {
3520                 ret = ret < 0 ? ret : -EIO;
3521                 end_extent_writepage(page, ret, start, page_end);
3522         }
3523         unlock_page(page);
3524         return ret;
3525
3526 done_unlocked:
3527         return 0;
3528 }
3529
3530 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3531 {
3532         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3533                        TASK_UNINTERRUPTIBLE);
3534 }
3535
3536 static noinline_for_stack int
3537 lock_extent_buffer_for_io(struct extent_buffer *eb,
3538                           struct btrfs_fs_info *fs_info,
3539                           struct extent_page_data *epd)
3540 {
3541         unsigned long i, num_pages;
3542         int flush = 0;
3543         int ret = 0;
3544
3545         if (!btrfs_try_tree_write_lock(eb)) {
3546                 flush = 1;
3547                 flush_write_bio(epd);
3548                 btrfs_tree_lock(eb);
3549         }
3550
3551         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3552                 btrfs_tree_unlock(eb);
3553                 if (!epd->sync_io)
3554                         return 0;
3555                 if (!flush) {
3556                         flush_write_bio(epd);
3557                         flush = 1;
3558                 }
3559                 while (1) {
3560                         wait_on_extent_buffer_writeback(eb);
3561                         btrfs_tree_lock(eb);
3562                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3563                                 break;
3564                         btrfs_tree_unlock(eb);
3565                 }
3566         }
3567
3568         /*
3569          * We need to do this to prevent races in people who check if the eb is
3570          * under IO since we can end up having no IO bits set for a short period
3571          * of time.
3572          */
3573         spin_lock(&eb->refs_lock);
3574         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3575                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3576                 spin_unlock(&eb->refs_lock);
3577                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3578                 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
3579                                          -eb->len,
3580                                          fs_info->dirty_metadata_batch);
3581                 ret = 1;
3582         } else {
3583                 spin_unlock(&eb->refs_lock);
3584         }
3585
3586         btrfs_tree_unlock(eb);
3587
3588         if (!ret)
3589                 return ret;
3590
3591         num_pages = num_extent_pages(eb->start, eb->len);
3592         for (i = 0; i < num_pages; i++) {
3593                 struct page *p = eb->pages[i];
3594
3595                 if (!trylock_page(p)) {
3596                         if (!flush) {
3597                                 flush_write_bio(epd);
3598                                 flush = 1;
3599                         }
3600                         lock_page(p);
3601                 }
3602         }
3603
3604         return ret;
3605 }
3606
3607 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3608 {
3609         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3610         smp_mb__after_atomic();
3611         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3612 }
3613
3614 static void set_btree_ioerr(struct page *page)
3615 {
3616         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3617
3618         SetPageError(page);
3619         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3620                 return;
3621
3622         /*
3623          * If writeback for a btree extent that doesn't belong to a log tree
3624          * failed, increment the counter transaction->eb_write_errors.
3625          * We do this because while the transaction is running and before it's
3626          * committing (when we call filemap_fdata[write|wait]_range against
3627          * the btree inode), we might have
3628          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3629          * returns an error or an error happens during writeback, when we're
3630          * committing the transaction we wouldn't know about it, since the pages
3631          * can be no longer dirty nor marked anymore for writeback (if a
3632          * subsequent modification to the extent buffer didn't happen before the
3633          * transaction commit), which makes filemap_fdata[write|wait]_range not
3634          * able to find the pages tagged with SetPageError at transaction
3635          * commit time. So if this happens we must abort the transaction,
3636          * otherwise we commit a super block with btree roots that point to
3637          * btree nodes/leafs whose content on disk is invalid - either garbage
3638          * or the content of some node/leaf from a past generation that got
3639          * cowed or deleted and is no longer valid.
3640          *
3641          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3642          * not be enough - we need to distinguish between log tree extents vs
3643          * non-log tree extents, and the next filemap_fdatawait_range() call
3644          * will catch and clear such errors in the mapping - and that call might
3645          * be from a log sync and not from a transaction commit. Also, checking
3646          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3647          * not done and would not be reliable - the eb might have been released
3648          * from memory and reading it back again means that flag would not be
3649          * set (since it's a runtime flag, not persisted on disk).
3650          *
3651          * Using the flags below in the btree inode also makes us achieve the
3652          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3653          * writeback for all dirty pages and before filemap_fdatawait_range()
3654          * is called, the writeback for all dirty pages had already finished
3655          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3656          * filemap_fdatawait_range() would return success, as it could not know
3657          * that writeback errors happened (the pages were no longer tagged for
3658          * writeback).
3659          */
3660         switch (eb->log_index) {
3661         case -1:
3662                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3663                 break;
3664         case 0:
3665                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3666                 break;
3667         case 1:
3668                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3669                 break;
3670         default:
3671                 BUG(); /* unexpected, logic error */
3672         }
3673 }
3674
3675 static void end_bio_extent_buffer_writepage(struct bio *bio)
3676 {
3677         struct bio_vec *bvec;
3678         struct extent_buffer *eb;
3679         int i, done;
3680
3681         ASSERT(!bio_flagged(bio, BIO_CLONED));
3682         bio_for_each_segment_all(bvec, bio, i) {
3683                 struct page *page = bvec->bv_page;
3684
3685                 eb = (struct extent_buffer *)page->private;
3686                 BUG_ON(!eb);
3687                 done = atomic_dec_and_test(&eb->io_pages);
3688
3689                 if (bio->bi_status ||
3690                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3691                         ClearPageUptodate(page);
3692                         set_btree_ioerr(page);
3693                 }
3694
3695                 end_page_writeback(page);
3696
3697                 if (!done)
3698                         continue;
3699
3700                 end_extent_buffer_writeback(eb);
3701         }
3702
3703         bio_put(bio);
3704 }
3705
3706 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3707                         struct btrfs_fs_info *fs_info,
3708                         struct writeback_control *wbc,
3709                         struct extent_page_data *epd)
3710 {
3711         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3712         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3713         u64 offset = eb->start;
3714         u32 nritems;
3715         unsigned long i, num_pages;
3716         unsigned long bio_flags = 0;
3717         unsigned long start, end;
3718         int write_flags = (epd->sync_io ? REQ_SYNC : 0) | REQ_META;
3719         int ret = 0;
3720
3721         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3722         num_pages = num_extent_pages(eb->start, eb->len);
3723         atomic_set(&eb->io_pages, num_pages);
3724         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3725                 bio_flags = EXTENT_BIO_TREE_LOG;
3726
3727         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3728         nritems = btrfs_header_nritems(eb);
3729         if (btrfs_header_level(eb) > 0) {
3730                 end = btrfs_node_key_ptr_offset(nritems);
3731
3732                 memzero_extent_buffer(eb, end, eb->len - end);
3733         } else {
3734                 /*
3735                  * leaf:
3736                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3737                  */
3738                 start = btrfs_item_nr_offset(nritems);
3739                 end = BTRFS_LEAF_DATA_OFFSET + leaf_data_end(fs_info, eb);
3740                 memzero_extent_buffer(eb, start, end - start);
3741         }
3742
3743         for (i = 0; i < num_pages; i++) {
3744                 struct page *p = eb->pages[i];
3745
3746                 clear_page_dirty_for_io(p);
3747                 set_page_writeback(p);
3748                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3749                                          p, offset >> 9, PAGE_SIZE, 0, bdev,
3750                                          &epd->bio,
3751                                          end_bio_extent_buffer_writepage,
3752                                          0, epd->bio_flags, bio_flags, false);
3753                 epd->bio_flags = bio_flags;
3754                 if (ret) {
3755                         set_btree_ioerr(p);
3756                         if (PageWriteback(p))
3757                                 end_page_writeback(p);
3758                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3759                                 end_extent_buffer_writeback(eb);
3760                         ret = -EIO;
3761                         break;
3762                 }
3763                 offset += PAGE_SIZE;
3764                 update_nr_written(wbc, 1);
3765                 unlock_page(p);
3766         }
3767
3768         if (unlikely(ret)) {
3769                 for (; i < num_pages; i++) {
3770                         struct page *p = eb->pages[i];
3771                         clear_page_dirty_for_io(p);
3772                         unlock_page(p);
3773                 }
3774         }
3775
3776         return ret;
3777 }
3778
3779 int btree_write_cache_pages(struct address_space *mapping,
3780                                    struct writeback_control *wbc)
3781 {
3782         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3783         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3784         struct extent_buffer *eb, *prev_eb = NULL;
3785         struct extent_page_data epd = {
3786                 .bio = NULL,
3787                 .tree = tree,
3788                 .extent_locked = 0,
3789                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3790                 .bio_flags = 0,
3791         };
3792         int ret = 0;
3793         int done = 0;
3794         int nr_to_write_done = 0;
3795         struct pagevec pvec;
3796         int nr_pages;
3797         pgoff_t index;
3798         pgoff_t end;            /* Inclusive */
3799         int scanned = 0;
3800         int tag;
3801
3802         pagevec_init(&pvec, 0);
3803         if (wbc->range_cyclic) {
3804                 index = mapping->writeback_index; /* Start from prev offset */
3805                 end = -1;
3806         } else {
3807                 index = wbc->range_start >> PAGE_SHIFT;
3808                 end = wbc->range_end >> PAGE_SHIFT;
3809                 scanned = 1;
3810         }
3811         if (wbc->sync_mode == WB_SYNC_ALL)
3812                 tag = PAGECACHE_TAG_TOWRITE;
3813         else
3814                 tag = PAGECACHE_TAG_DIRTY;
3815 retry:
3816         if (wbc->sync_mode == WB_SYNC_ALL)
3817                 tag_pages_for_writeback(mapping, index, end);
3818         while (!done && !nr_to_write_done && (index <= end) &&
3819                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3820                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3821                 unsigned i;
3822
3823                 scanned = 1;
3824                 for (i = 0; i < nr_pages; i++) {
3825                         struct page *page = pvec.pages[i];
3826
3827                         if (!PagePrivate(page))
3828                                 continue;
3829
3830                         if (!wbc->range_cyclic && page->index > end) {
3831                                 done = 1;
3832                                 break;
3833                         }
3834
3835                         spin_lock(&mapping->private_lock);
3836                         if (!PagePrivate(page)) {
3837                                 spin_unlock(&mapping->private_lock);
3838                                 continue;
3839                         }
3840
3841                         eb = (struct extent_buffer *)page->private;
3842
3843                         /*
3844                          * Shouldn't happen and normally this would be a BUG_ON
3845                          * but no sense in crashing the users box for something
3846                          * we can survive anyway.
3847                          */
3848                         if (WARN_ON(!eb)) {
3849                                 spin_unlock(&mapping->private_lock);
3850                                 continue;
3851                         }
3852
3853                         if (eb == prev_eb) {
3854                                 spin_unlock(&mapping->private_lock);
3855                                 continue;
3856                         }
3857
3858                         ret = atomic_inc_not_zero(&eb->refs);
3859                         spin_unlock(&mapping->private_lock);
3860                         if (!ret)
3861                                 continue;
3862
3863                         prev_eb = eb;
3864                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3865                         if (!ret) {
3866                                 free_extent_buffer(eb);
3867                                 continue;
3868                         }
3869
3870                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3871                         if (ret) {
3872                                 done = 1;
3873                                 free_extent_buffer(eb);
3874                                 break;
3875                         }
3876                         free_extent_buffer(eb);
3877
3878                         /*
3879                          * the filesystem may choose to bump up nr_to_write.
3880                          * We have to make sure to honor the new nr_to_write
3881                          * at any time
3882                          */
3883                         nr_to_write_done = wbc->nr_to_write <= 0;
3884                 }
3885                 pagevec_release(&pvec);
3886                 cond_resched();
3887         }
3888         if (!scanned && !done) {
3889                 /*
3890                  * We hit the last page and there is more work to be done: wrap
3891                  * back to the start of the file
3892                  */
3893                 scanned = 1;
3894                 index = 0;
3895                 goto retry;
3896         }
3897         flush_write_bio(&epd);
3898         return ret;
3899 }
3900
3901 /**
3902  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3903  * @mapping: address space structure to write
3904  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3905  * @writepage: function called for each page
3906  * @data: data passed to writepage function
3907  *
3908  * If a page is already under I/O, write_cache_pages() skips it, even
3909  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3910  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3911  * and msync() need to guarantee that all the data which was dirty at the time
3912  * the call was made get new I/O started against them.  If wbc->sync_mode is
3913  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3914  * existing IO to complete.
3915  */
3916 static int extent_write_cache_pages(struct address_space *mapping,
3917                              struct writeback_control *wbc,
3918                              writepage_t writepage, void *data,
3919                              void (*flush_fn)(void *))
3920 {
3921         struct inode *inode = mapping->host;
3922         int ret = 0;
3923         int done = 0;
3924         int nr_to_write_done = 0;
3925         struct pagevec pvec;
3926         int nr_pages;
3927         pgoff_t index;
3928         pgoff_t end;            /* Inclusive */
3929         pgoff_t done_index;
3930         int range_whole = 0;
3931         int scanned = 0;
3932         int tag;
3933
3934         /*
3935          * We have to hold onto the inode so that ordered extents can do their
3936          * work when the IO finishes.  The alternative to this is failing to add
3937          * an ordered extent if the igrab() fails there and that is a huge pain
3938          * to deal with, so instead just hold onto the inode throughout the
3939          * writepages operation.  If it fails here we are freeing up the inode
3940          * anyway and we'd rather not waste our time writing out stuff that is
3941          * going to be truncated anyway.
3942          */
3943         if (!igrab(inode))
3944                 return 0;
3945
3946         pagevec_init(&pvec, 0);
3947         if (wbc->range_cyclic) {
3948                 index = mapping->writeback_index; /* Start from prev offset */
3949                 end = -1;
3950         } else {
3951                 index = wbc->range_start >> PAGE_SHIFT;
3952                 end = wbc->range_end >> PAGE_SHIFT;
3953                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3954                         range_whole = 1;
3955                 scanned = 1;
3956         }
3957         if (wbc->sync_mode == WB_SYNC_ALL)
3958                 tag = PAGECACHE_TAG_TOWRITE;
3959         else
3960                 tag = PAGECACHE_TAG_DIRTY;
3961 retry:
3962         if (wbc->sync_mode == WB_SYNC_ALL)
3963                 tag_pages_for_writeback(mapping, index, end);
3964         done_index = index;
3965         while (!done && !nr_to_write_done && (index <= end) &&
3966                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3967                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3968                 unsigned i;
3969
3970                 scanned = 1;
3971                 for (i = 0; i < nr_pages; i++) {
3972                         struct page *page = pvec.pages[i];
3973
3974                         done_index = page->index;
3975                         /*
3976                          * At this point we hold neither mapping->tree_lock nor
3977                          * lock on the page itself: the page may be truncated or
3978                          * invalidated (changing page->mapping to NULL), or even
3979                          * swizzled back from swapper_space to tmpfs file
3980                          * mapping
3981                          */
3982                         if (!trylock_page(page)) {
3983                                 flush_fn(data);
3984                                 lock_page(page);
3985                         }
3986
3987                         if (unlikely(page->mapping != mapping)) {
3988                                 unlock_page(page);
3989                                 continue;
3990                         }
3991
3992                         if (!wbc->range_cyclic && page->index > end) {
3993                                 done = 1;
3994                                 unlock_page(page);
3995                                 continue;
3996                         }
3997
3998                         if (wbc->sync_mode != WB_SYNC_NONE) {
3999                                 if (PageWriteback(page))
4000                                         flush_fn(data);
4001                                 wait_on_page_writeback(page);
4002                         }
4003
4004                         if (PageWriteback(page) ||
4005                             !clear_page_dirty_for_io(page)) {
4006                                 unlock_page(page);
4007                                 continue;
4008                         }
4009
4010                         ret = (*writepage)(page, wbc, data);
4011
4012                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4013                                 unlock_page(page);
4014                                 ret = 0;
4015                         }
4016                         if (ret < 0) {
4017                                 /*
4018                                  * done_index is set past this page,
4019                                  * so media errors will not choke
4020                                  * background writeout for the entire
4021                                  * file. This has consequences for
4022                                  * range_cyclic semantics (ie. it may
4023                                  * not be suitable for data integrity
4024                                  * writeout).
4025                                  */
4026                                 done_index = page->index + 1;
4027                                 done = 1;
4028                                 break;
4029                         }
4030
4031                         /*
4032                          * the filesystem may choose to bump up nr_to_write.
4033                          * We have to make sure to honor the new nr_to_write
4034                          * at any time
4035                          */
4036                         nr_to_write_done = wbc->nr_to_write <= 0;
4037                 }
4038                 pagevec_release(&pvec);
4039                 cond_resched();
4040         }
4041         if (!scanned && !done) {
4042                 /*
4043                  * We hit the last page and there is more work to be done: wrap
4044                  * back to the start of the file
4045                  */
4046                 scanned = 1;
4047                 index = 0;
4048                 goto retry;
4049         }
4050
4051         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4052                 mapping->writeback_index = done_index;
4053
4054         btrfs_add_delayed_iput(inode);
4055         return ret;
4056 }
4057
4058 static void flush_epd_write_bio(struct extent_page_data *epd)
4059 {
4060         if (epd->bio) {
4061                 int ret;
4062
4063                 bio_set_op_attrs(epd->bio, REQ_OP_WRITE,
4064                                  epd->sync_io ? REQ_SYNC : 0);
4065
4066                 ret = submit_one_bio(epd->bio, 0, epd->bio_flags);
4067                 BUG_ON(ret < 0); /* -ENOMEM */
4068                 epd->bio = NULL;
4069         }
4070 }
4071
4072 static noinline void flush_write_bio(void *data)
4073 {
4074         struct extent_page_data *epd = data;
4075         flush_epd_write_bio(epd);
4076 }
4077
4078 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4079                           get_extent_t *get_extent,
4080                           struct writeback_control *wbc)
4081 {
4082         int ret;
4083         struct extent_page_data epd = {
4084                 .bio = NULL,
4085                 .tree = tree,
4086                 .get_extent = get_extent,
4087                 .extent_locked = 0,
4088                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4089                 .bio_flags = 0,
4090         };
4091
4092         ret = __extent_writepage(page, wbc, &epd);
4093
4094         flush_epd_write_bio(&epd);
4095         return ret;
4096 }
4097
4098 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4099                               u64 start, u64 end, get_extent_t *get_extent,
4100                               int mode)
4101 {
4102         int ret = 0;
4103         struct address_space *mapping = inode->i_mapping;
4104         struct page *page;
4105         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4106                 PAGE_SHIFT;
4107
4108         struct extent_page_data epd = {
4109                 .bio = NULL,
4110                 .tree = tree,
4111                 .get_extent = get_extent,
4112                 .extent_locked = 1,
4113                 .sync_io = mode == WB_SYNC_ALL,
4114                 .bio_flags = 0,
4115         };
4116         struct writeback_control wbc_writepages = {
4117                 .sync_mode      = mode,
4118                 .nr_to_write    = nr_pages * 2,
4119                 .range_start    = start,
4120                 .range_end      = end + 1,
4121         };
4122
4123         while (start <= end) {
4124                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4125                 if (clear_page_dirty_for_io(page))
4126                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4127                 else {
4128                         if (tree->ops && tree->ops->writepage_end_io_hook)
4129                                 tree->ops->writepage_end_io_hook(page, start,
4130                                                  start + PAGE_SIZE - 1,
4131                                                  NULL, 1);
4132                         unlock_page(page);
4133                 }
4134                 put_page(page);
4135                 start += PAGE_SIZE;
4136         }
4137
4138         flush_epd_write_bio(&epd);
4139         return ret;
4140 }
4141
4142 int extent_writepages(struct extent_io_tree *tree,
4143                       struct address_space *mapping,
4144                       get_extent_t *get_extent,
4145                       struct writeback_control *wbc)
4146 {
4147         int ret = 0;
4148         struct extent_page_data epd = {
4149                 .bio = NULL,
4150                 .tree = tree,
4151                 .get_extent = get_extent,
4152                 .extent_locked = 0,
4153                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4154                 .bio_flags = 0,
4155         };
4156
4157         ret = extent_write_cache_pages(mapping, wbc, __extent_writepage, &epd,
4158                                        flush_write_bio);
4159         flush_epd_write_bio(&epd);
4160         return ret;
4161 }
4162
4163 int extent_readpages(struct extent_io_tree *tree,
4164                      struct address_space *mapping,
4165                      struct list_head *pages, unsigned nr_pages,
4166                      get_extent_t get_extent)
4167 {
4168         struct bio *bio = NULL;
4169         unsigned page_idx;
4170         unsigned long bio_flags = 0;
4171         struct page *pagepool[16];
4172         struct page *page;
4173         struct extent_map *em_cached = NULL;
4174         int nr = 0;
4175         u64 prev_em_start = (u64)-1;
4176
4177         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4178                 page = list_entry(pages->prev, struct page, lru);
4179
4180                 prefetchw(&page->flags);
4181                 list_del(&page->lru);
4182                 if (add_to_page_cache_lru(page, mapping,
4183                                         page->index,
4184                                         readahead_gfp_mask(mapping))) {
4185                         put_page(page);
4186                         continue;
4187                 }
4188
4189                 pagepool[nr++] = page;
4190                 if (nr < ARRAY_SIZE(pagepool))
4191                         continue;
4192                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4193                                    &bio, 0, &bio_flags, &prev_em_start);
4194                 nr = 0;
4195         }
4196         if (nr)
4197                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4198                                    &bio, 0, &bio_flags, &prev_em_start);
4199
4200         if (em_cached)
4201                 free_extent_map(em_cached);
4202
4203         BUG_ON(!list_empty(pages));
4204         if (bio)
4205                 return submit_one_bio(bio, 0, bio_flags);
4206         return 0;
4207 }
4208
4209 /*
4210  * basic invalidatepage code, this waits on any locked or writeback
4211  * ranges corresponding to the page, and then deletes any extent state
4212  * records from the tree
4213  */
4214 int extent_invalidatepage(struct extent_io_tree *tree,
4215                           struct page *page, unsigned long offset)
4216 {
4217         struct extent_state *cached_state = NULL;
4218         u64 start = page_offset(page);
4219         u64 end = start + PAGE_SIZE - 1;
4220         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4221
4222         start += ALIGN(offset, blocksize);
4223         if (start > end)
4224                 return 0;
4225
4226         lock_extent_bits(tree, start, end, &cached_state);
4227         wait_on_page_writeback(page);
4228         clear_extent_bit(tree, start, end,
4229                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4230                          EXTENT_DO_ACCOUNTING,
4231                          1, 1, &cached_state, GFP_NOFS);
4232         return 0;
4233 }
4234
4235 /*
4236  * a helper for releasepage, this tests for areas of the page that
4237  * are locked or under IO and drops the related state bits if it is safe
4238  * to drop the page.
4239  */
4240 static int try_release_extent_state(struct extent_map_tree *map,
4241                                     struct extent_io_tree *tree,
4242                                     struct page *page, gfp_t mask)
4243 {
4244         u64 start = page_offset(page);
4245         u64 end = start + PAGE_SIZE - 1;
4246         int ret = 1;
4247
4248         if (test_range_bit(tree, start, end,
4249                            EXTENT_IOBITS, 0, NULL))
4250                 ret = 0;
4251         else {
4252                 /*
4253                  * at this point we can safely clear everything except the
4254                  * locked bit and the nodatasum bit
4255                  */
4256                 ret = clear_extent_bit(tree, start, end,
4257                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4258                                  0, 0, NULL, mask);
4259
4260                 /* if clear_extent_bit failed for enomem reasons,
4261                  * we can't allow the release to continue.
4262                  */
4263                 if (ret < 0)
4264                         ret = 0;
4265                 else
4266                         ret = 1;
4267         }
4268         return ret;
4269 }
4270
4271 /*
4272  * a helper for releasepage.  As long as there are no locked extents
4273  * in the range corresponding to the page, both state records and extent
4274  * map records are removed
4275  */
4276 int try_release_extent_mapping(struct extent_map_tree *map,
4277                                struct extent_io_tree *tree, struct page *page,
4278                                gfp_t mask)
4279 {
4280         struct extent_map *em;
4281         u64 start = page_offset(page);
4282         u64 end = start + PAGE_SIZE - 1;
4283
4284         if (gfpflags_allow_blocking(mask) &&
4285             page->mapping->host->i_size > SZ_16M) {
4286                 u64 len;
4287                 while (start <= end) {
4288                         len = end - start + 1;
4289                         write_lock(&map->lock);
4290                         em = lookup_extent_mapping(map, start, len);
4291                         if (!em) {
4292                                 write_unlock(&map->lock);
4293                                 break;
4294                         }
4295                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4296                             em->start != start) {
4297                                 write_unlock(&map->lock);
4298                                 free_extent_map(em);
4299                                 break;
4300                         }
4301                         if (!test_range_bit(tree, em->start,
4302                                             extent_map_end(em) - 1,
4303                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4304                                             0, NULL)) {
4305                                 remove_extent_mapping(map, em);
4306                                 /* once for the rb tree */
4307                                 free_extent_map(em);
4308                         }
4309                         start = extent_map_end(em);
4310                         write_unlock(&map->lock);
4311
4312                         /* once for us */
4313                         free_extent_map(em);
4314                 }
4315         }
4316         return try_release_extent_state(map, tree, page, mask);
4317 }
4318
4319 /*
4320  * helper function for fiemap, which doesn't want to see any holes.
4321  * This maps until we find something past 'last'
4322  */
4323 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4324                                                 u64 offset,
4325                                                 u64 last,
4326                                                 get_extent_t *get_extent)
4327 {
4328         u64 sectorsize = btrfs_inode_sectorsize(inode);
4329         struct extent_map *em;
4330         u64 len;
4331
4332         if (offset >= last)
4333                 return NULL;
4334
4335         while (1) {
4336                 len = last - offset;
4337                 if (len == 0)
4338                         break;
4339                 len = ALIGN(len, sectorsize);
4340                 em = get_extent(BTRFS_I(inode), NULL, 0, offset, len, 0);
4341                 if (IS_ERR_OR_NULL(em))
4342                         return em;
4343
4344                 /* if this isn't a hole return it */
4345                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4346                     em->block_start != EXTENT_MAP_HOLE) {
4347                         return em;
4348                 }
4349
4350                 /* this is a hole, advance to the next extent */
4351                 offset = extent_map_end(em);
4352                 free_extent_map(em);
4353                 if (offset >= last)
4354                         break;
4355         }
4356         return NULL;
4357 }
4358
4359 /*
4360  * To cache previous fiemap extent
4361  *
4362  * Will be used for merging fiemap extent
4363  */
4364 struct fiemap_cache {
4365         u64 offset;
4366         u64 phys;
4367         u64 len;
4368         u32 flags;
4369         bool cached;
4370 };
4371
4372 /*
4373  * Helper to submit fiemap extent.
4374  *
4375  * Will try to merge current fiemap extent specified by @offset, @phys,
4376  * @len and @flags with cached one.
4377  * And only when we fails to merge, cached one will be submitted as
4378  * fiemap extent.
4379  *
4380  * Return value is the same as fiemap_fill_next_extent().
4381  */
4382 static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
4383                                 struct fiemap_cache *cache,
4384                                 u64 offset, u64 phys, u64 len, u32 flags)
4385 {
4386         int ret = 0;
4387
4388         if (!cache->cached)
4389                 goto assign;
4390
4391         /*
4392          * Sanity check, extent_fiemap() should have ensured that new
4393          * fiemap extent won't overlap with cahced one.
4394          * Not recoverable.
4395          *
4396          * NOTE: Physical address can overlap, due to compression
4397          */
4398         if (cache->offset + cache->len > offset) {
4399                 WARN_ON(1);
4400                 return -EINVAL;
4401         }
4402
4403         /*
4404          * Only merges fiemap extents if
4405          * 1) Their logical addresses are continuous
4406          *
4407          * 2) Their physical addresses are continuous
4408          *    So truly compressed (physical size smaller than logical size)
4409          *    extents won't get merged with each other
4410          *
4411          * 3) Share same flags except FIEMAP_EXTENT_LAST
4412          *    So regular extent won't get merged with prealloc extent
4413          */
4414         if (cache->offset + cache->len  == offset &&
4415             cache->phys + cache->len == phys  &&
4416             (cache->flags & ~FIEMAP_EXTENT_LAST) ==
4417                         (flags & ~FIEMAP_EXTENT_LAST)) {
4418                 cache->len += len;
4419                 cache->flags |= flags;
4420                 goto try_submit_last;
4421         }
4422
4423         /* Not mergeable, need to submit cached one */
4424         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4425                                       cache->len, cache->flags);
4426         cache->cached = false;
4427         if (ret)
4428                 return ret;
4429 assign:
4430         cache->cached = true;
4431         cache->offset = offset;
4432         cache->phys = phys;
4433         cache->len = len;
4434         cache->flags = flags;
4435 try_submit_last:
4436         if (cache->flags & FIEMAP_EXTENT_LAST) {
4437                 ret = fiemap_fill_next_extent(fieinfo, cache->offset,
4438                                 cache->phys, cache->len, cache->flags);
4439                 cache->cached = false;
4440         }
4441         return ret;
4442 }
4443
4444 /*
4445  * Emit last fiemap cache
4446  *
4447  * The last fiemap cache may still be cached in the following case:
4448  * 0                  4k                    8k
4449  * |<- Fiemap range ->|
4450  * |<------------  First extent ----------->|
4451  *
4452  * In this case, the first extent range will be cached but not emitted.
4453  * So we must emit it before ending extent_fiemap().
4454  */
4455 static int emit_last_fiemap_cache(struct btrfs_fs_info *fs_info,
4456                                   struct fiemap_extent_info *fieinfo,
4457                                   struct fiemap_cache *cache)
4458 {
4459         int ret;
4460
4461         if (!cache->cached)
4462                 return 0;
4463
4464         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4465                                       cache->len, cache->flags);
4466         cache->cached = false;
4467         if (ret > 0)
4468                 ret = 0;
4469         return ret;
4470 }
4471
4472 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4473                 __u64 start, __u64 len, get_extent_t *get_extent)
4474 {
4475         int ret = 0;
4476         u64 off = start;
4477         u64 max = start + len;
4478         u32 flags = 0;
4479         u32 found_type;
4480         u64 last;
4481         u64 last_for_get_extent = 0;
4482         u64 disko = 0;
4483         u64 isize = i_size_read(inode);
4484         struct btrfs_key found_key;
4485         struct extent_map *em = NULL;
4486         struct extent_state *cached_state = NULL;
4487         struct btrfs_path *path;
4488         struct btrfs_root *root = BTRFS_I(inode)->root;
4489         struct fiemap_cache cache = { 0 };
4490         int end = 0;
4491         u64 em_start = 0;
4492         u64 em_len = 0;
4493         u64 em_end = 0;
4494
4495         if (len == 0)
4496                 return -EINVAL;
4497
4498         path = btrfs_alloc_path();
4499         if (!path)
4500                 return -ENOMEM;
4501         path->leave_spinning = 1;
4502
4503         start = round_down(start, btrfs_inode_sectorsize(inode));
4504         len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
4505
4506         /*
4507          * lookup the last file extent.  We're not using i_size here
4508          * because there might be preallocation past i_size
4509          */
4510         ret = btrfs_lookup_file_extent(NULL, root, path,
4511                         btrfs_ino(BTRFS_I(inode)), -1, 0);
4512         if (ret < 0) {
4513                 btrfs_free_path(path);
4514                 return ret;
4515         } else {
4516                 WARN_ON(!ret);
4517                 if (ret == 1)
4518                         ret = 0;
4519         }
4520
4521         path->slots[0]--;
4522         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4523         found_type = found_key.type;
4524
4525         /* No extents, but there might be delalloc bits */
4526         if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) ||
4527             found_type != BTRFS_EXTENT_DATA_KEY) {
4528                 /* have to trust i_size as the end */
4529                 last = (u64)-1;
4530                 last_for_get_extent = isize;
4531         } else {
4532                 /*
4533                  * remember the start of the last extent.  There are a
4534                  * bunch of different factors that go into the length of the
4535                  * extent, so its much less complex to remember where it started
4536                  */
4537                 last = found_key.offset;
4538                 last_for_get_extent = last + 1;
4539         }
4540         btrfs_release_path(path);
4541
4542         /*
4543          * we might have some extents allocated but more delalloc past those
4544          * extents.  so, we trust isize unless the start of the last extent is
4545          * beyond isize
4546          */
4547         if (last < isize) {
4548                 last = (u64)-1;
4549                 last_for_get_extent = isize;
4550         }
4551
4552         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4553                          &cached_state);
4554
4555         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4556                                    get_extent);
4557         if (!em)
4558                 goto out;
4559         if (IS_ERR(em)) {
4560                 ret = PTR_ERR(em);
4561                 goto out;
4562         }
4563
4564         while (!end) {
4565                 u64 offset_in_extent = 0;
4566
4567                 /* break if the extent we found is outside the range */
4568                 if (em->start >= max || extent_map_end(em) < off)
4569                         break;
4570
4571                 /*
4572                  * get_extent may return an extent that starts before our
4573                  * requested range.  We have to make sure the ranges
4574                  * we return to fiemap always move forward and don't
4575                  * overlap, so adjust the offsets here
4576                  */
4577                 em_start = max(em->start, off);
4578
4579                 /*
4580                  * record the offset from the start of the extent
4581                  * for adjusting the disk offset below.  Only do this if the
4582                  * extent isn't compressed since our in ram offset may be past
4583                  * what we have actually allocated on disk.
4584                  */
4585                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4586                         offset_in_extent = em_start - em->start;
4587                 em_end = extent_map_end(em);
4588                 em_len = em_end - em_start;
4589                 disko = 0;
4590                 flags = 0;
4591
4592                 /*
4593                  * bump off for our next call to get_extent
4594                  */
4595                 off = extent_map_end(em);
4596                 if (off >= max)
4597                         end = 1;
4598
4599                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4600                         end = 1;
4601                         flags |= FIEMAP_EXTENT_LAST;
4602                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4603                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4604                                   FIEMAP_EXTENT_NOT_ALIGNED);
4605                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4606                         flags |= (FIEMAP_EXTENT_DELALLOC |
4607                                   FIEMAP_EXTENT_UNKNOWN);
4608                 } else if (fieinfo->fi_extents_max) {
4609                         struct btrfs_trans_handle *trans;
4610
4611                         u64 bytenr = em->block_start -
4612                                 (em->start - em->orig_start);
4613
4614                         disko = em->block_start + offset_in_extent;
4615
4616                         /*
4617                          * We need a trans handle to get delayed refs
4618                          */
4619                         trans = btrfs_join_transaction(root);
4620                         /*
4621                          * It's OK if we can't start a trans we can still check
4622                          * from commit_root
4623                          */
4624                         if (IS_ERR(trans))
4625                                 trans = NULL;
4626
4627                         /*
4628                          * As btrfs supports shared space, this information
4629                          * can be exported to userspace tools via
4630                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4631                          * then we're just getting a count and we can skip the
4632                          * lookup stuff.
4633                          */
4634                         ret = btrfs_check_shared(trans, root->fs_info,
4635                                         root->objectid,
4636                                         btrfs_ino(BTRFS_I(inode)), bytenr);
4637                         if (trans)
4638                                 btrfs_end_transaction(trans);
4639                         if (ret < 0)
4640                                 goto out_free;
4641                         if (ret)
4642                                 flags |= FIEMAP_EXTENT_SHARED;
4643                         ret = 0;
4644                 }
4645                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4646                         flags |= FIEMAP_EXTENT_ENCODED;
4647                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4648                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4649
4650                 free_extent_map(em);
4651                 em = NULL;
4652                 if ((em_start >= last) || em_len == (u64)-1 ||
4653                    (last == (u64)-1 && isize <= em_end)) {
4654                         flags |= FIEMAP_EXTENT_LAST;
4655                         end = 1;
4656                 }
4657
4658                 /* now scan forward to see if this is really the last extent. */
4659                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4660                                            get_extent);
4661                 if (IS_ERR(em)) {
4662                         ret = PTR_ERR(em);
4663                         goto out;
4664                 }
4665                 if (!em) {
4666                         flags |= FIEMAP_EXTENT_LAST;
4667                         end = 1;
4668                 }
4669                 ret = emit_fiemap_extent(fieinfo, &cache, em_start, disko,
4670                                            em_len, flags);
4671                 if (ret) {
4672                         if (ret == 1)
4673                                 ret = 0;
4674                         goto out_free;
4675                 }
4676         }
4677 out_free:
4678         if (!ret)
4679                 ret = emit_last_fiemap_cache(root->fs_info, fieinfo, &cache);
4680         free_extent_map(em);
4681 out:
4682         btrfs_free_path(path);
4683         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4684                              &cached_state, GFP_NOFS);
4685         return ret;
4686 }
4687
4688 static void __free_extent_buffer(struct extent_buffer *eb)
4689 {
4690         btrfs_leak_debug_del(&eb->leak_list);
4691         kmem_cache_free(extent_buffer_cache, eb);
4692 }
4693
4694 int extent_buffer_under_io(struct extent_buffer *eb)
4695 {
4696         return (atomic_read(&eb->io_pages) ||
4697                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4698                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4699 }
4700
4701 /*
4702  * Helper for releasing extent buffer page.
4703  */
4704 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4705 {
4706         unsigned long index;
4707         struct page *page;
4708         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4709
4710         BUG_ON(extent_buffer_under_io(eb));
4711
4712         index = num_extent_pages(eb->start, eb->len);
4713         if (index == 0)
4714                 return;
4715
4716         do {
4717                 index--;
4718                 page = eb->pages[index];
4719                 if (!page)
4720                         continue;
4721                 if (mapped)
4722                         spin_lock(&page->mapping->private_lock);
4723                 /*
4724                  * We do this since we'll remove the pages after we've
4725                  * removed the eb from the radix tree, so we could race
4726                  * and have this page now attached to the new eb.  So
4727                  * only clear page_private if it's still connected to
4728                  * this eb.
4729                  */
4730                 if (PagePrivate(page) &&
4731                     page->private == (unsigned long)eb) {
4732                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4733                         BUG_ON(PageDirty(page));
4734                         BUG_ON(PageWriteback(page));
4735                         /*
4736                          * We need to make sure we haven't be attached
4737                          * to a new eb.
4738                          */
4739                         ClearPagePrivate(page);
4740                         set_page_private(page, 0);
4741                         /* One for the page private */
4742                         put_page(page);
4743                 }
4744
4745                 if (mapped)
4746                         spin_unlock(&page->mapping->private_lock);
4747
4748                 /* One for when we allocated the page */
4749                 put_page(page);
4750         } while (index != 0);
4751 }
4752
4753 /*
4754  * Helper for releasing the extent buffer.
4755  */
4756 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4757 {
4758         btrfs_release_extent_buffer_page(eb);
4759         __free_extent_buffer(eb);
4760 }
4761
4762 static struct extent_buffer *
4763 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4764                       unsigned long len)
4765 {
4766         struct extent_buffer *eb = NULL;
4767
4768         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4769         eb->start = start;
4770         eb->len = len;
4771         eb->fs_info = fs_info;
4772         eb->bflags = 0;
4773         rwlock_init(&eb->lock);
4774         atomic_set(&eb->write_locks, 0);
4775         atomic_set(&eb->read_locks, 0);
4776         atomic_set(&eb->blocking_readers, 0);
4777         atomic_set(&eb->blocking_writers, 0);
4778         atomic_set(&eb->spinning_readers, 0);
4779         atomic_set(&eb->spinning_writers, 0);
4780         eb->lock_nested = 0;
4781         init_waitqueue_head(&eb->write_lock_wq);
4782         init_waitqueue_head(&eb->read_lock_wq);
4783
4784         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4785
4786         spin_lock_init(&eb->refs_lock);
4787         atomic_set(&eb->refs, 1);
4788         atomic_set(&eb->io_pages, 0);
4789
4790         /*
4791          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4792          */
4793         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4794                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4795         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4796
4797         return eb;
4798 }
4799
4800 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4801 {
4802         unsigned long i;
4803         struct page *p;
4804         struct extent_buffer *new;
4805         unsigned long num_pages = num_extent_pages(src->start, src->len);
4806
4807         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4808         if (new == NULL)
4809                 return NULL;
4810
4811         for (i = 0; i < num_pages; i++) {
4812                 p = alloc_page(GFP_NOFS);
4813                 if (!p) {
4814                         btrfs_release_extent_buffer(new);
4815                         return NULL;
4816                 }
4817                 attach_extent_buffer_page(new, p);
4818                 WARN_ON(PageDirty(p));
4819                 SetPageUptodate(p);
4820                 new->pages[i] = p;
4821                 copy_page(page_address(p), page_address(src->pages[i]));
4822         }
4823
4824         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4825         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4826
4827         return new;
4828 }
4829
4830 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4831                                                   u64 start, unsigned long len)
4832 {
4833         struct extent_buffer *eb;
4834         unsigned long num_pages;
4835         unsigned long i;
4836
4837         num_pages = num_extent_pages(start, len);
4838
4839         eb = __alloc_extent_buffer(fs_info, start, len);
4840         if (!eb)
4841                 return NULL;
4842
4843         for (i = 0; i < num_pages; i++) {
4844                 eb->pages[i] = alloc_page(GFP_NOFS);
4845                 if (!eb->pages[i])
4846                         goto err;
4847         }
4848         set_extent_buffer_uptodate(eb);
4849         btrfs_set_header_nritems(eb, 0);
4850         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4851
4852         return eb;
4853 err:
4854         for (; i > 0; i--)
4855                 __free_page(eb->pages[i - 1]);
4856         __free_extent_buffer(eb);
4857         return NULL;
4858 }
4859
4860 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4861                                                 u64 start)
4862 {
4863         return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
4864 }
4865
4866 static void check_buffer_tree_ref(struct extent_buffer *eb)
4867 {
4868         int refs;
4869         /* the ref bit is tricky.  We have to make sure it is set
4870          * if we have the buffer dirty.   Otherwise the
4871          * code to free a buffer can end up dropping a dirty
4872          * page
4873          *
4874          * Once the ref bit is set, it won't go away while the
4875          * buffer is dirty or in writeback, and it also won't
4876          * go away while we have the reference count on the
4877          * eb bumped.
4878          *
4879          * We can't just set the ref bit without bumping the
4880          * ref on the eb because free_extent_buffer might
4881          * see the ref bit and try to clear it.  If this happens
4882          * free_extent_buffer might end up dropping our original
4883          * ref by mistake and freeing the page before we are able
4884          * to add one more ref.
4885          *
4886          * So bump the ref count first, then set the bit.  If someone
4887          * beat us to it, drop the ref we added.
4888          */
4889         refs = atomic_read(&eb->refs);
4890         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4891                 return;
4892
4893         spin_lock(&eb->refs_lock);
4894         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4895                 atomic_inc(&eb->refs);
4896         spin_unlock(&eb->refs_lock);
4897 }
4898
4899 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4900                 struct page *accessed)
4901 {
4902         unsigned long num_pages, i;
4903
4904         check_buffer_tree_ref(eb);
4905
4906         num_pages = num_extent_pages(eb->start, eb->len);
4907         for (i = 0; i < num_pages; i++) {
4908                 struct page *p = eb->pages[i];
4909
4910                 if (p != accessed)
4911                         mark_page_accessed(p);
4912         }
4913 }
4914
4915 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4916                                          u64 start)
4917 {
4918         struct extent_buffer *eb;
4919
4920         rcu_read_lock();
4921         eb = radix_tree_lookup(&fs_info->buffer_radix,
4922                                start >> PAGE_SHIFT);
4923         if (eb && atomic_inc_not_zero(&eb->refs)) {
4924                 rcu_read_unlock();
4925                 /*
4926                  * Lock our eb's refs_lock to avoid races with
4927                  * free_extent_buffer. When we get our eb it might be flagged
4928                  * with EXTENT_BUFFER_STALE and another task running
4929                  * free_extent_buffer might have seen that flag set,
4930                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4931                  * writeback flags not set) and it's still in the tree (flag
4932                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4933                  * of decrementing the extent buffer's reference count twice.
4934                  * So here we could race and increment the eb's reference count,
4935                  * clear its stale flag, mark it as dirty and drop our reference
4936                  * before the other task finishes executing free_extent_buffer,
4937                  * which would later result in an attempt to free an extent
4938                  * buffer that is dirty.
4939                  */
4940                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4941                         spin_lock(&eb->refs_lock);
4942                         spin_unlock(&eb->refs_lock);
4943                 }
4944                 mark_extent_buffer_accessed(eb, NULL);
4945                 return eb;
4946         }
4947         rcu_read_unlock();
4948
4949         return NULL;
4950 }
4951
4952 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4953 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4954                                         u64 start)
4955 {
4956         struct extent_buffer *eb, *exists = NULL;
4957         int ret;
4958
4959         eb = find_extent_buffer(fs_info, start);
4960         if (eb)
4961                 return eb;
4962         eb = alloc_dummy_extent_buffer(fs_info, start);
4963         if (!eb)
4964                 return NULL;
4965         eb->fs_info = fs_info;
4966 again:
4967         ret = radix_tree_preload(GFP_NOFS);
4968         if (ret)
4969                 goto free_eb;
4970         spin_lock(&fs_info->buffer_lock);
4971         ret = radix_tree_insert(&fs_info->buffer_radix,
4972                                 start >> PAGE_SHIFT, eb);
4973         spin_unlock(&fs_info->buffer_lock);
4974         radix_tree_preload_end();
4975         if (ret == -EEXIST) {
4976                 exists = find_extent_buffer(fs_info, start);
4977                 if (exists)
4978                         goto free_eb;
4979                 else
4980                         goto again;
4981         }
4982         check_buffer_tree_ref(eb);
4983         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4984
4985         /*
4986          * We will free dummy extent buffer's if they come into
4987          * free_extent_buffer with a ref count of 2, but if we are using this we
4988          * want the buffers to stay in memory until we're done with them, so
4989          * bump the ref count again.
4990          */
4991         atomic_inc(&eb->refs);
4992         return eb;
4993 free_eb:
4994         btrfs_release_extent_buffer(eb);
4995         return exists;
4996 }
4997 #endif
4998
4999 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
5000                                           u64 start)
5001 {
5002         unsigned long len = fs_info->nodesize;
5003         unsigned long num_pages = num_extent_pages(start, len);
5004         unsigned long i;
5005         unsigned long index = start >> PAGE_SHIFT;
5006         struct extent_buffer *eb;
5007         struct extent_buffer *exists = NULL;
5008         struct page *p;
5009         struct address_space *mapping = fs_info->btree_inode->i_mapping;
5010         int uptodate = 1;
5011         int ret;
5012
5013         if (!IS_ALIGNED(start, fs_info->sectorsize)) {
5014                 btrfs_err(fs_info, "bad tree block start %llu", start);
5015                 return ERR_PTR(-EINVAL);
5016         }
5017
5018         eb = find_extent_buffer(fs_info, start);
5019         if (eb)
5020                 return eb;
5021
5022         eb = __alloc_extent_buffer(fs_info, start, len);
5023         if (!eb)
5024                 return ERR_PTR(-ENOMEM);
5025
5026         for (i = 0; i < num_pages; i++, index++) {
5027                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
5028                 if (!p) {
5029                         exists = ERR_PTR(-ENOMEM);
5030                         goto free_eb;
5031                 }
5032
5033                 spin_lock(&mapping->private_lock);
5034                 if (PagePrivate(p)) {
5035                         /*
5036                          * We could have already allocated an eb for this page
5037                          * and attached one so lets see if we can get a ref on
5038                          * the existing eb, and if we can we know it's good and
5039                          * we can just return that one, else we know we can just
5040                          * overwrite page->private.
5041                          */
5042                         exists = (struct extent_buffer *)p->private;
5043                         if (atomic_inc_not_zero(&exists->refs)) {
5044                                 spin_unlock(&mapping->private_lock);
5045                                 unlock_page(p);
5046                                 put_page(p);
5047                                 mark_extent_buffer_accessed(exists, p);
5048                                 goto free_eb;
5049                         }
5050                         exists = NULL;
5051
5052                         /*
5053                          * Do this so attach doesn't complain and we need to
5054                          * drop the ref the old guy had.
5055                          */
5056                         ClearPagePrivate(p);
5057                         WARN_ON(PageDirty(p));
5058                         put_page(p);
5059                 }
5060                 attach_extent_buffer_page(eb, p);
5061                 spin_unlock(&mapping->private_lock);
5062                 WARN_ON(PageDirty(p));
5063                 eb->pages[i] = p;
5064                 if (!PageUptodate(p))
5065                         uptodate = 0;
5066
5067                 /*
5068                  * see below about how we avoid a nasty race with release page
5069                  * and why we unlock later
5070                  */
5071         }
5072         if (uptodate)
5073                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5074 again:
5075         ret = radix_tree_preload(GFP_NOFS);
5076         if (ret) {
5077                 exists = ERR_PTR(ret);
5078                 goto free_eb;
5079         }
5080
5081         spin_lock(&fs_info->buffer_lock);
5082         ret = radix_tree_insert(&fs_info->buffer_radix,
5083                                 start >> PAGE_SHIFT, eb);
5084         spin_unlock(&fs_info->buffer_lock);
5085         radix_tree_preload_end();
5086         if (ret == -EEXIST) {
5087                 exists = find_extent_buffer(fs_info, start);
5088                 if (exists)
5089                         goto free_eb;
5090                 else
5091                         goto again;
5092         }
5093         /* add one reference for the tree */
5094         check_buffer_tree_ref(eb);
5095         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5096
5097         /*
5098          * there is a race where release page may have
5099          * tried to find this extent buffer in the radix
5100          * but failed.  It will tell the VM it is safe to
5101          * reclaim the, and it will clear the page private bit.
5102          * We must make sure to set the page private bit properly
5103          * after the extent buffer is in the radix tree so
5104          * it doesn't get lost
5105          */
5106         SetPageChecked(eb->pages[0]);
5107         for (i = 1; i < num_pages; i++) {
5108                 p = eb->pages[i];
5109                 ClearPageChecked(p);
5110                 unlock_page(p);
5111         }
5112         unlock_page(eb->pages[0]);
5113         return eb;
5114
5115 free_eb:
5116         WARN_ON(!atomic_dec_and_test(&eb->refs));
5117         for (i = 0; i < num_pages; i++) {
5118                 if (eb->pages[i])
5119                         unlock_page(eb->pages[i]);
5120         }
5121
5122         btrfs_release_extent_buffer(eb);
5123         return exists;
5124 }
5125
5126 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5127 {
5128         struct extent_buffer *eb =
5129                         container_of(head, struct extent_buffer, rcu_head);
5130
5131         __free_extent_buffer(eb);
5132 }
5133
5134 /* Expects to have eb->eb_lock already held */
5135 static int release_extent_buffer(struct extent_buffer *eb)
5136 {
5137         WARN_ON(atomic_read(&eb->refs) == 0);
5138         if (atomic_dec_and_test(&eb->refs)) {
5139                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5140                         struct btrfs_fs_info *fs_info = eb->fs_info;
5141
5142                         spin_unlock(&eb->refs_lock);
5143
5144                         spin_lock(&fs_info->buffer_lock);
5145                         radix_tree_delete(&fs_info->buffer_radix,
5146                                           eb->start >> PAGE_SHIFT);
5147                         spin_unlock(&fs_info->buffer_lock);
5148                 } else {
5149                         spin_unlock(&eb->refs_lock);
5150                 }
5151
5152                 /* Should be safe to release our pages at this point */
5153                 btrfs_release_extent_buffer_page(eb);
5154 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5155                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5156                         __free_extent_buffer(eb);
5157                         return 1;
5158                 }
5159 #endif
5160                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5161                 return 1;
5162         }
5163         spin_unlock(&eb->refs_lock);
5164
5165         return 0;
5166 }
5167
5168 void free_extent_buffer(struct extent_buffer *eb)
5169 {
5170         int refs;
5171         int old;
5172         if (!eb)
5173                 return;
5174
5175         while (1) {
5176                 refs = atomic_read(&eb->refs);
5177                 if (refs <= 3)
5178                         break;
5179                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5180                 if (old == refs)
5181                         return;
5182         }
5183
5184         spin_lock(&eb->refs_lock);
5185         if (atomic_read(&eb->refs) == 2 &&
5186             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5187                 atomic_dec(&eb->refs);
5188
5189         if (atomic_read(&eb->refs) == 2 &&
5190             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5191             !extent_buffer_under_io(eb) &&
5192             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5193                 atomic_dec(&eb->refs);
5194
5195         /*
5196          * I know this is terrible, but it's temporary until we stop tracking
5197          * the uptodate bits and such for the extent buffers.
5198          */
5199         release_extent_buffer(eb);
5200 }
5201
5202 void free_extent_buffer_stale(struct extent_buffer *eb)
5203 {
5204         if (!eb)
5205                 return;
5206
5207         spin_lock(&eb->refs_lock);
5208         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5209
5210         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5211             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5212                 atomic_dec(&eb->refs);
5213         release_extent_buffer(eb);
5214 }
5215
5216 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5217 {
5218         unsigned long i;
5219         unsigned long num_pages;
5220         struct page *page;
5221
5222         num_pages = num_extent_pages(eb->start, eb->len);
5223
5224         for (i = 0; i < num_pages; i++) {
5225                 page = eb->pages[i];
5226                 if (!PageDirty(page))
5227                         continue;
5228
5229                 lock_page(page);
5230                 WARN_ON(!PagePrivate(page));
5231
5232                 clear_page_dirty_for_io(page);
5233                 spin_lock_irq(&page->mapping->tree_lock);
5234                 if (!PageDirty(page)) {
5235                         radix_tree_tag_clear(&page->mapping->page_tree,
5236                                                 page_index(page),
5237                                                 PAGECACHE_TAG_DIRTY);
5238                 }
5239                 spin_unlock_irq(&page->mapping->tree_lock);
5240                 ClearPageError(page);
5241                 unlock_page(page);
5242         }
5243         WARN_ON(atomic_read(&eb->refs) == 0);
5244 }
5245
5246 int set_extent_buffer_dirty(struct extent_buffer *eb)
5247 {
5248         unsigned long i;
5249         unsigned long num_pages;
5250         int was_dirty = 0;
5251
5252         check_buffer_tree_ref(eb);
5253
5254         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5255
5256         num_pages = num_extent_pages(eb->start, eb->len);
5257         WARN_ON(atomic_read(&eb->refs) == 0);
5258         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5259
5260         for (i = 0; i < num_pages; i++)
5261                 set_page_dirty(eb->pages[i]);
5262         return was_dirty;
5263 }
5264
5265 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5266 {
5267         unsigned long i;
5268         struct page *page;
5269         unsigned long num_pages;
5270
5271         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5272         num_pages = num_extent_pages(eb->start, eb->len);
5273         for (i = 0; i < num_pages; i++) {
5274                 page = eb->pages[i];
5275                 if (page)
5276                         ClearPageUptodate(page);
5277         }
5278 }
5279
5280 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5281 {
5282         unsigned long i;
5283         struct page *page;
5284         unsigned long num_pages;
5285
5286         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5287         num_pages = num_extent_pages(eb->start, eb->len);
5288         for (i = 0; i < num_pages; i++) {
5289                 page = eb->pages[i];
5290                 SetPageUptodate(page);
5291         }
5292 }
5293
5294 int extent_buffer_uptodate(struct extent_buffer *eb)
5295 {
5296         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5297 }
5298
5299 int read_extent_buffer_pages(struct extent_io_tree *tree,
5300                              struct extent_buffer *eb, int wait,
5301                              get_extent_t *get_extent, int mirror_num)
5302 {
5303         unsigned long i;
5304         struct page *page;
5305         int err;
5306         int ret = 0;
5307         int locked_pages = 0;
5308         int all_uptodate = 1;
5309         unsigned long num_pages;
5310         unsigned long num_reads = 0;
5311         struct bio *bio = NULL;
5312         unsigned long bio_flags = 0;
5313
5314         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5315                 return 0;
5316
5317         num_pages = num_extent_pages(eb->start, eb->len);
5318         for (i = 0; i < num_pages; i++) {
5319                 page = eb->pages[i];
5320                 if (wait == WAIT_NONE) {
5321                         if (!trylock_page(page))
5322                                 goto unlock_exit;
5323                 } else {
5324                         lock_page(page);
5325                 }
5326                 locked_pages++;
5327         }
5328         /*
5329          * We need to firstly lock all pages to make sure that
5330          * the uptodate bit of our pages won't be affected by
5331          * clear_extent_buffer_uptodate().
5332          */
5333         for (i = 0; i < num_pages; i++) {
5334                 page = eb->pages[i];
5335                 if (!PageUptodate(page)) {
5336                         num_reads++;
5337                         all_uptodate = 0;
5338                 }
5339         }
5340
5341         if (all_uptodate) {
5342                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5343                 goto unlock_exit;
5344         }
5345
5346         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5347         eb->read_mirror = 0;
5348         atomic_set(&eb->io_pages, num_reads);
5349         for (i = 0; i < num_pages; i++) {
5350                 page = eb->pages[i];
5351
5352                 if (!PageUptodate(page)) {
5353                         if (ret) {
5354                                 atomic_dec(&eb->io_pages);
5355                                 unlock_page(page);
5356                                 continue;
5357                         }
5358
5359                         ClearPageError(page);
5360                         err = __extent_read_full_page(tree, page,
5361                                                       get_extent, &bio,
5362                                                       mirror_num, &bio_flags,
5363                                                       REQ_META);
5364                         if (err) {
5365                                 ret = err;
5366                                 /*
5367                                  * We use &bio in above __extent_read_full_page,
5368                                  * so we ensure that if it returns error, the
5369                                  * current page fails to add itself to bio and
5370                                  * it's been unlocked.
5371                                  *
5372                                  * We must dec io_pages by ourselves.
5373                                  */
5374                                 atomic_dec(&eb->io_pages);
5375                         }
5376                 } else {
5377                         unlock_page(page);
5378                 }
5379         }
5380
5381         if (bio) {
5382                 err = submit_one_bio(bio, mirror_num, bio_flags);
5383                 if (err)
5384                         return err;
5385         }
5386
5387         if (ret || wait != WAIT_COMPLETE)
5388                 return ret;
5389
5390         for (i = 0; i < num_pages; i++) {
5391                 page = eb->pages[i];
5392                 wait_on_page_locked(page);
5393                 if (!PageUptodate(page))
5394                         ret = -EIO;
5395         }
5396
5397         return ret;
5398
5399 unlock_exit:
5400         while (locked_pages > 0) {
5401                 locked_pages--;
5402                 page = eb->pages[locked_pages];
5403                 unlock_page(page);
5404         }
5405         return ret;
5406 }
5407
5408 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5409                         unsigned long start,
5410                         unsigned long len)
5411 {
5412         size_t cur;
5413         size_t offset;
5414         struct page *page;
5415         char *kaddr;
5416         char *dst = (char *)dstv;
5417         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5418         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5419
5420         WARN_ON(start > eb->len);
5421         WARN_ON(start + len > eb->start + eb->len);
5422
5423         offset = (start_offset + start) & (PAGE_SIZE - 1);
5424
5425         while (len > 0) {
5426                 page = eb->pages[i];
5427
5428                 cur = min(len, (PAGE_SIZE - offset));
5429                 kaddr = page_address(page);
5430                 memcpy(dst, kaddr + offset, cur);
5431
5432                 dst += cur;
5433                 len -= cur;
5434                 offset = 0;
5435                 i++;
5436         }
5437 }
5438
5439 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5440                         unsigned long start,
5441                         unsigned long len)
5442 {
5443         size_t cur;
5444         size_t offset;
5445         struct page *page;
5446         char *kaddr;
5447         char __user *dst = (char __user *)dstv;
5448         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5449         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5450         int ret = 0;
5451
5452         WARN_ON(start > eb->len);
5453         WARN_ON(start + len > eb->start + eb->len);
5454
5455         offset = (start_offset + start) & (PAGE_SIZE - 1);
5456
5457         while (len > 0) {
5458                 page = eb->pages[i];
5459
5460                 cur = min(len, (PAGE_SIZE - offset));
5461                 kaddr = page_address(page);
5462                 if (copy_to_user(dst, kaddr + offset, cur)) {
5463                         ret = -EFAULT;
5464                         break;
5465                 }
5466
5467                 dst += cur;
5468                 len -= cur;
5469                 offset = 0;
5470                 i++;
5471         }
5472
5473         return ret;
5474 }
5475
5476 /*
5477  * return 0 if the item is found within a page.
5478  * return 1 if the item spans two pages.
5479  * return -EINVAL otherwise.
5480  */
5481 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5482                                unsigned long min_len, char **map,
5483                                unsigned long *map_start,
5484                                unsigned long *map_len)
5485 {
5486         size_t offset = start & (PAGE_SIZE - 1);
5487         char *kaddr;
5488         struct page *p;
5489         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5490         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5491         unsigned long end_i = (start_offset + start + min_len - 1) >>
5492                 PAGE_SHIFT;
5493
5494         if (i != end_i)
5495                 return 1;
5496
5497         if (i == 0) {
5498                 offset = start_offset;
5499                 *map_start = 0;
5500         } else {
5501                 offset = 0;
5502                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5503         }
5504
5505         if (start + min_len > eb->len) {
5506                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5507                        eb->start, eb->len, start, min_len);
5508                 return -EINVAL;
5509         }
5510
5511         p = eb->pages[i];
5512         kaddr = page_address(p);
5513         *map = kaddr + offset;
5514         *map_len = PAGE_SIZE - offset;
5515         return 0;
5516 }
5517
5518 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5519                           unsigned long start,
5520                           unsigned long len)
5521 {
5522         size_t cur;
5523         size_t offset;
5524         struct page *page;
5525         char *kaddr;
5526         char *ptr = (char *)ptrv;
5527         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5528         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5529         int ret = 0;
5530
5531         WARN_ON(start > eb->len);
5532         WARN_ON(start + len > eb->start + eb->len);
5533
5534         offset = (start_offset + start) & (PAGE_SIZE - 1);
5535
5536         while (len > 0) {
5537                 page = eb->pages[i];
5538
5539                 cur = min(len, (PAGE_SIZE - offset));
5540
5541                 kaddr = page_address(page);
5542                 ret = memcmp(ptr, kaddr + offset, cur);
5543                 if (ret)
5544                         break;
5545
5546                 ptr += cur;
5547                 len -= cur;
5548                 offset = 0;
5549                 i++;
5550         }
5551         return ret;
5552 }
5553
5554 void write_extent_buffer_chunk_tree_uuid(struct extent_buffer *eb,
5555                 const void *srcv)
5556 {
5557         char *kaddr;
5558
5559         WARN_ON(!PageUptodate(eb->pages[0]));
5560         kaddr = page_address(eb->pages[0]);
5561         memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv,
5562                         BTRFS_FSID_SIZE);
5563 }
5564
5565 void write_extent_buffer_fsid(struct extent_buffer *eb, const void *srcv)
5566 {
5567         char *kaddr;
5568
5569         WARN_ON(!PageUptodate(eb->pages[0]));
5570         kaddr = page_address(eb->pages[0]);
5571         memcpy(kaddr + offsetof(struct btrfs_header, fsid), srcv,
5572                         BTRFS_FSID_SIZE);
5573 }
5574
5575 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5576                          unsigned long start, unsigned long len)
5577 {
5578         size_t cur;
5579         size_t offset;
5580         struct page *page;
5581         char *kaddr;
5582         char *src = (char *)srcv;
5583         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5584         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5585
5586         WARN_ON(start > eb->len);
5587         WARN_ON(start + len > eb->start + eb->len);
5588
5589         offset = (start_offset + start) & (PAGE_SIZE - 1);
5590
5591         while (len > 0) {
5592                 page = eb->pages[i];
5593                 WARN_ON(!PageUptodate(page));
5594
5595                 cur = min(len, PAGE_SIZE - offset);
5596                 kaddr = page_address(page);
5597                 memcpy(kaddr + offset, src, cur);
5598
5599                 src += cur;
5600                 len -= cur;
5601                 offset = 0;
5602                 i++;
5603         }
5604 }
5605
5606 void memzero_extent_buffer(struct extent_buffer *eb, unsigned long start,
5607                 unsigned long len)
5608 {
5609         size_t cur;
5610         size_t offset;
5611         struct page *page;
5612         char *kaddr;
5613         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5614         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5615
5616         WARN_ON(start > eb->len);
5617         WARN_ON(start + len > eb->start + eb->len);
5618
5619         offset = (start_offset + start) & (PAGE_SIZE - 1);
5620
5621         while (len > 0) {
5622                 page = eb->pages[i];
5623                 WARN_ON(!PageUptodate(page));
5624
5625                 cur = min(len, PAGE_SIZE - offset);
5626                 kaddr = page_address(page);
5627                 memset(kaddr + offset, 0, cur);
5628
5629                 len -= cur;
5630                 offset = 0;
5631                 i++;
5632         }
5633 }
5634
5635 void copy_extent_buffer_full(struct extent_buffer *dst,
5636                              struct extent_buffer *src)
5637 {
5638         int i;
5639         unsigned num_pages;
5640
5641         ASSERT(dst->len == src->len);
5642
5643         num_pages = num_extent_pages(dst->start, dst->len);
5644         for (i = 0; i < num_pages; i++)
5645                 copy_page(page_address(dst->pages[i]),
5646                                 page_address(src->pages[i]));
5647 }
5648
5649 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5650                         unsigned long dst_offset, unsigned long src_offset,
5651                         unsigned long len)
5652 {
5653         u64 dst_len = dst->len;
5654         size_t cur;
5655         size_t offset;
5656         struct page *page;
5657         char *kaddr;
5658         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5659         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5660
5661         WARN_ON(src->len != dst_len);
5662
5663         offset = (start_offset + dst_offset) &
5664                 (PAGE_SIZE - 1);
5665
5666         while (len > 0) {
5667                 page = dst->pages[i];
5668                 WARN_ON(!PageUptodate(page));
5669
5670                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5671
5672                 kaddr = page_address(page);
5673                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5674
5675                 src_offset += cur;
5676                 len -= cur;
5677                 offset = 0;
5678                 i++;
5679         }
5680 }
5681
5682 void le_bitmap_set(u8 *map, unsigned int start, int len)
5683 {
5684         u8 *p = map + BIT_BYTE(start);
5685         const unsigned int size = start + len;
5686         int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5687         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);
5688
5689         while (len - bits_to_set >= 0) {
5690                 *p |= mask_to_set;
5691                 len -= bits_to_set;
5692                 bits_to_set = BITS_PER_BYTE;
5693                 mask_to_set = ~0;
5694                 p++;
5695         }
5696         if (len) {
5697                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5698                 *p |= mask_to_set;
5699         }
5700 }
5701
5702 void le_bitmap_clear(u8 *map, unsigned int start, int len)
5703 {
5704         u8 *p = map + BIT_BYTE(start);
5705         const unsigned int size = start + len;
5706         int bits_to_clear = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5707         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(start);
5708
5709         while (len - bits_to_clear >= 0) {
5710                 *p &= ~mask_to_clear;
5711                 len -= bits_to_clear;
5712                 bits_to_clear = BITS_PER_BYTE;
5713                 mask_to_clear = ~0;
5714                 p++;
5715         }
5716         if (len) {
5717                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5718                 *p &= ~mask_to_clear;
5719         }
5720 }
5721
5722 /*
5723  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5724  * given bit number
5725  * @eb: the extent buffer
5726  * @start: offset of the bitmap item in the extent buffer
5727  * @nr: bit number
5728  * @page_index: return index of the page in the extent buffer that contains the
5729  * given bit number
5730  * @page_offset: return offset into the page given by page_index
5731  *
5732  * This helper hides the ugliness of finding the byte in an extent buffer which
5733  * contains a given bit.
5734  */
5735 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5736                                     unsigned long start, unsigned long nr,
5737                                     unsigned long *page_index,
5738                                     size_t *page_offset)
5739 {
5740         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5741         size_t byte_offset = BIT_BYTE(nr);
5742         size_t offset;
5743
5744         /*
5745          * The byte we want is the offset of the extent buffer + the offset of
5746          * the bitmap item in the extent buffer + the offset of the byte in the
5747          * bitmap item.
5748          */
5749         offset = start_offset + start + byte_offset;
5750
5751         *page_index = offset >> PAGE_SHIFT;
5752         *page_offset = offset & (PAGE_SIZE - 1);
5753 }
5754
5755 /**
5756  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5757  * @eb: the extent buffer
5758  * @start: offset of the bitmap item in the extent buffer
5759  * @nr: bit number to test
5760  */
5761 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5762                            unsigned long nr)
5763 {
5764         u8 *kaddr;
5765         struct page *page;
5766         unsigned long i;
5767         size_t offset;
5768
5769         eb_bitmap_offset(eb, start, nr, &i, &offset);
5770         page = eb->pages[i];
5771         WARN_ON(!PageUptodate(page));
5772         kaddr = page_address(page);
5773         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5774 }
5775
5776 /**
5777  * extent_buffer_bitmap_set - set an area of a bitmap
5778  * @eb: the extent buffer
5779  * @start: offset of the bitmap item in the extent buffer
5780  * @pos: bit number of the first bit
5781  * @len: number of bits to set
5782  */
5783 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5784                               unsigned long pos, unsigned long len)
5785 {
5786         u8 *kaddr;
5787         struct page *page;
5788         unsigned long i;
5789         size_t offset;
5790         const unsigned int size = pos + len;
5791         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5792         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5793
5794         eb_bitmap_offset(eb, start, pos, &i, &offset);
5795         page = eb->pages[i];
5796         WARN_ON(!PageUptodate(page));
5797         kaddr = page_address(page);
5798
5799         while (len >= bits_to_set) {
5800                 kaddr[offset] |= mask_to_set;
5801                 len -= bits_to_set;
5802                 bits_to_set = BITS_PER_BYTE;
5803                 mask_to_set = ~0;
5804                 if (++offset >= PAGE_SIZE && len > 0) {
5805                         offset = 0;
5806                         page = eb->pages[++i];
5807                         WARN_ON(!PageUptodate(page));
5808                         kaddr = page_address(page);
5809                 }
5810         }
5811         if (len) {
5812                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5813                 kaddr[offset] |= mask_to_set;
5814         }
5815 }
5816
5817
5818 /**
5819  * extent_buffer_bitmap_clear - clear an area of a bitmap
5820  * @eb: the extent buffer
5821  * @start: offset of the bitmap item in the extent buffer
5822  * @pos: bit number of the first bit
5823  * @len: number of bits to clear
5824  */
5825 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5826                                 unsigned long pos, unsigned long len)
5827 {
5828         u8 *kaddr;
5829         struct page *page;
5830         unsigned long i;
5831         size_t offset;
5832         const unsigned int size = pos + len;
5833         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5834         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5835
5836         eb_bitmap_offset(eb, start, pos, &i, &offset);
5837         page = eb->pages[i];
5838         WARN_ON(!PageUptodate(page));
5839         kaddr = page_address(page);
5840
5841         while (len >= bits_to_clear) {
5842                 kaddr[offset] &= ~mask_to_clear;
5843                 len -= bits_to_clear;
5844                 bits_to_clear = BITS_PER_BYTE;
5845                 mask_to_clear = ~0;
5846                 if (++offset >= PAGE_SIZE && len > 0) {
5847                         offset = 0;
5848                         page = eb->pages[++i];
5849                         WARN_ON(!PageUptodate(page));
5850                         kaddr = page_address(page);
5851                 }
5852         }
5853         if (len) {
5854                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5855                 kaddr[offset] &= ~mask_to_clear;
5856         }
5857 }
5858
5859 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5860 {
5861         unsigned long distance = (src > dst) ? src - dst : dst - src;
5862         return distance < len;
5863 }
5864
5865 static void copy_pages(struct page *dst_page, struct page *src_page,
5866                        unsigned long dst_off, unsigned long src_off,
5867                        unsigned long len)
5868 {
5869         char *dst_kaddr = page_address(dst_page);
5870         char *src_kaddr;
5871         int must_memmove = 0;
5872
5873         if (dst_page != src_page) {
5874                 src_kaddr = page_address(src_page);
5875         } else {
5876                 src_kaddr = dst_kaddr;
5877                 if (areas_overlap(src_off, dst_off, len))
5878                         must_memmove = 1;
5879         }
5880
5881         if (must_memmove)
5882                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5883         else
5884                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5885 }
5886
5887 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5888                            unsigned long src_offset, unsigned long len)
5889 {
5890         struct btrfs_fs_info *fs_info = dst->fs_info;
5891         size_t cur;
5892         size_t dst_off_in_page;
5893         size_t src_off_in_page;
5894         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5895         unsigned long dst_i;
5896         unsigned long src_i;
5897
5898         if (src_offset + len > dst->len) {
5899                 btrfs_err(fs_info,
5900                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5901                          src_offset, len, dst->len);
5902                 BUG_ON(1);
5903         }
5904         if (dst_offset + len > dst->len) {
5905                 btrfs_err(fs_info,
5906                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5907                          dst_offset, len, dst->len);
5908                 BUG_ON(1);
5909         }
5910
5911         while (len > 0) {
5912                 dst_off_in_page = (start_offset + dst_offset) &
5913                         (PAGE_SIZE - 1);
5914                 src_off_in_page = (start_offset + src_offset) &
5915                         (PAGE_SIZE - 1);
5916
5917                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5918                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5919
5920                 cur = min(len, (unsigned long)(PAGE_SIZE -
5921                                                src_off_in_page));
5922                 cur = min_t(unsigned long, cur,
5923                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5924
5925                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5926                            dst_off_in_page, src_off_in_page, cur);
5927
5928                 src_offset += cur;
5929                 dst_offset += cur;
5930                 len -= cur;
5931         }
5932 }
5933
5934 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5935                            unsigned long src_offset, unsigned long len)
5936 {
5937         struct btrfs_fs_info *fs_info = dst->fs_info;
5938         size_t cur;
5939         size_t dst_off_in_page;
5940         size_t src_off_in_page;
5941         unsigned long dst_end = dst_offset + len - 1;
5942         unsigned long src_end = src_offset + len - 1;
5943         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5944         unsigned long dst_i;
5945         unsigned long src_i;
5946
5947         if (src_offset + len > dst->len) {
5948                 btrfs_err(fs_info,
5949                           "memmove bogus src_offset %lu move len %lu len %lu",
5950                           src_offset, len, dst->len);
5951                 BUG_ON(1);
5952         }
5953         if (dst_offset + len > dst->len) {
5954                 btrfs_err(fs_info,
5955                           "memmove bogus dst_offset %lu move len %lu len %lu",
5956                           dst_offset, len, dst->len);
5957                 BUG_ON(1);
5958         }
5959         if (dst_offset < src_offset) {
5960                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5961                 return;
5962         }
5963         while (len > 0) {
5964                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5965                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5966
5967                 dst_off_in_page = (start_offset + dst_end) &
5968                         (PAGE_SIZE - 1);
5969                 src_off_in_page = (start_offset + src_end) &
5970                         (PAGE_SIZE - 1);
5971
5972                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5973                 cur = min(cur, dst_off_in_page + 1);
5974                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5975                            dst_off_in_page - cur + 1,
5976                            src_off_in_page - cur + 1, cur);
5977
5978                 dst_end -= cur;
5979                 src_end -= cur;
5980                 len -= cur;
5981         }
5982 }
5983
5984 int try_release_extent_buffer(struct page *page)
5985 {
5986         struct extent_buffer *eb;
5987
5988         /*
5989          * We need to make sure nobody is attaching this page to an eb right
5990          * now.
5991          */
5992         spin_lock(&page->mapping->private_lock);
5993         if (!PagePrivate(page)) {
5994                 spin_unlock(&page->mapping->private_lock);
5995                 return 1;
5996         }
5997
5998         eb = (struct extent_buffer *)page->private;
5999         BUG_ON(!eb);
6000
6001         /*
6002          * This is a little awful but should be ok, we need to make sure that
6003          * the eb doesn't disappear out from under us while we're looking at
6004          * this page.
6005          */
6006         spin_lock(&eb->refs_lock);
6007         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
6008                 spin_unlock(&eb->refs_lock);
6009                 spin_unlock(&page->mapping->private_lock);
6010                 return 0;
6011         }
6012         spin_unlock(&page->mapping->private_lock);
6013
6014         /*
6015          * If tree ref isn't set then we know the ref on this eb is a real ref,
6016          * so just return, this page will likely be freed soon anyway.
6017          */
6018         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
6019                 spin_unlock(&eb->refs_lock);
6020                 return 0;
6021         }
6022
6023         return release_extent_buffer(eb);
6024 }