fs/btrfs/extent_io.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/bitops.h>
   3 #include <linux/slab.h>
   4 #include <linux/bio.h>
   5 #include <linux/mm.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/page-flags.h>
   8 #include <linux/spinlock.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/swap.h>
  11 #include <linux/writeback.h>
  12 #include <linux/pagevec.h>
  13 #include <linux/prefetch.h>
  14 #include <linux/cleancache.h>
  15 #include "extent_io.h"
  16 #include "extent_map.h"
  17 #include "ctree.h"
  18 #include "btrfs_inode.h"
  19 #include "volumes.h"
  20 #include "check-integrity.h"
  21 #include "locking.h"
  22 #include "rcu-string.h"
  23 #include "backref.h"
  24 #include "disk-io.h"
  25
  26 static struct kmem_cache *extent_state_cache;
  27 static struct kmem_cache *extent_buffer_cache;
  28 static struct bio_set *btrfs_bioset;
  29
  30 static inline bool extent_state_in_tree(const struct extent_state *state)
  31 {
  32         return !RB_EMPTY_NODE(&state->rb_node);
  33 }
  34
  35 #ifdef CONFIG_BTRFS_DEBUG
  36 static LIST_HEAD(buffers);
  37 static LIST_HEAD(states);
  38
  39 static DEFINE_SPINLOCK(leak_lock);
  40
  41 static inline
  42 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  43 {
  44         unsigned long flags;
  45
  46         spin_lock_irqsave(&leak_lock, flags);
  47         list_add(new, head);
  48         spin_unlock_irqrestore(&leak_lock, flags);
  49 }
  50
  51 static inline
  52 void btrfs_leak_debug_del(struct list_head *entry)
  53 {
  54         unsigned long flags;
  55
  56         spin_lock_irqsave(&leak_lock, flags);
  57         list_del(entry);
  58         spin_unlock_irqrestore(&leak_lock, flags);
  59 }
  60
  61 static inline
  62 void btrfs_leak_debug_check(void)
  63 {
  64         struct extent_state *state;
  65         struct extent_buffer *eb;
  66
  67         while (!list_empty(&states)) {
  68                 state = list_entry(states.next, struct extent_state, leak_list);
  69                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  70                        state->start, state->end, state->state,
  71                        extent_state_in_tree(state),
  72                        refcount_read(&state->refs));
  73                 list_del(&state->leak_list);
  74                 kmem_cache_free(extent_state_cache, state);
  75         }
  76
  77         while (!list_empty(&buffers)) {
  78                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  79                 pr_err("BTRFS: buffer leak start %llu len %lu refs %d\n",
  80                        eb->start, eb->len, atomic_read(&eb->refs));
  81                 list_del(&eb->leak_list);
  82                 kmem_cache_free(extent_buffer_cache, eb);
  83         }
  84 }
  85
  86 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  87         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  88 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  89                 struct extent_io_tree *tree, u64 start, u64 end)
  90 {
  91         if (tree->ops && tree->ops->check_extent_io_range)
  92                 tree->ops->check_extent_io_range(tree->private_data, caller,
  93                                                  start, end);
  94 }
  95 #else
  96 #define btrfs_leak_debug_add(new, head) do {} while (0)
  97 #define btrfs_leak_debug_del(entry)     do {} while (0)
  98 #define btrfs_leak_debug_check()        do {} while (0)
  99 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 100 #endif
 101
 102 #define BUFFER_LRU_MAX 64
 103
 104 struct tree_entry {
 105         u64 start;
 106         u64 end;
 107         struct rb_node rb_node;
 108 };
 109
 110 struct extent_page_data {
 111         struct bio *bio;
 112         struct extent_io_tree *tree;
 113         /* tells writepage not to lock the state bits for this range
 114          * it still does the unlocking
 115          */
 116         unsigned int extent_locked:1;
 117
 118         /* tells the submit_bio code to use REQ_SYNC */
 119         unsigned int sync_io:1;
 120 };
 121
 122 static void add_extent_changeset(struct extent_state *state, unsigned bits,
 123                                  struct extent_changeset *changeset,
 124                                  int set)
 125 {
 126         int ret;
 127
 128         if (!changeset)
 129                 return;
 130         if (set && (state->state & bits) == bits)
 131                 return;
 132         if (!set && (state->state & bits) == 0)
 133                 return;
 134         changeset->bytes_changed += state->end - state->start + 1;
 135         ret = ulist_add(&changeset->range_changed, state->start, state->end,
 136                         GFP_ATOMIC);
 137         /* ENOMEM */
 138         BUG_ON(ret < 0);
 139 }
 140
 141 static noinline void flush_write_bio(void *data);
 142 static inline struct btrfs_fs_info *
 143 tree_fs_info(struct extent_io_tree *tree)
 144 {
 145         if (tree->ops)
 146                 return tree->ops->tree_fs_info(tree->private_data);
 147         return NULL;
 148 }
 149
 150 int __init extent_io_init(void)
 151 {
 152         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 153                         sizeof(struct extent_state), 0,
 154                         SLAB_MEM_SPREAD, NULL);
 155         if (!extent_state_cache)
 156                 return -ENOMEM;
 157
 158         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 159                         sizeof(struct extent_buffer), 0,
 160                         SLAB_MEM_SPREAD, NULL);
 161         if (!extent_buffer_cache)
 162                 goto free_state_cache;
 163
 164         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 165                                      offsetof(struct btrfs_io_bio, bio),
 166                                      BIOSET_NEED_BVECS);
 167         if (!btrfs_bioset)
 168                 goto free_buffer_cache;
 169
 170         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 171                 goto free_bioset;
 172
 173         return 0;
 174
 175 free_bioset:
 176         bioset_free(btrfs_bioset);
 177         btrfs_bioset = NULL;
 178
 179 free_buffer_cache:
 180         kmem_cache_destroy(extent_buffer_cache);
 181         extent_buffer_cache = NULL;
 182
 183 free_state_cache:
 184         kmem_cache_destroy(extent_state_cache);
 185         extent_state_cache = NULL;
 186         return -ENOMEM;
 187 }
 188
 189 void extent_io_exit(void)
 190 {
 191         btrfs_leak_debug_check();
 192
 193         /*
 194          * Make sure all delayed rcu free are flushed before we
 195          * destroy caches.
 196          */
 197         rcu_barrier();
 198         kmem_cache_destroy(extent_state_cache);
 199         kmem_cache_destroy(extent_buffer_cache);
 200         if (btrfs_bioset)
 201                 bioset_free(btrfs_bioset);
 202 }
 203
 204 void extent_io_tree_init(struct extent_io_tree *tree,
 205                          void *private_data)
 206 {
 207         tree->state = RB_ROOT;
 208         tree->ops = NULL;
 209         tree->dirty_bytes = 0;
 210         spin_lock_init(&tree->lock);
 211         tree->private_data = private_data;
 212 }
 213
 214 static struct extent_state *alloc_extent_state(gfp_t mask)
 215 {
 216         struct extent_state *state;
 217
 218         /*
 219          * The given mask might be not appropriate for the slab allocator,
 220          * drop the unsupported bits
 221          */
 222         mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
 223         state = kmem_cache_alloc(extent_state_cache, mask);
 224         if (!state)
 225                 return state;
 226         state->state = 0;
 227         state->failrec = NULL;
 228         RB_CLEAR_NODE(&state->rb_node);
 229         btrfs_leak_debug_add(&state->leak_list, &states);
 230         refcount_set(&state->refs, 1);
 231         init_waitqueue_head(&state->wq);
 232         trace_alloc_extent_state(state, mask, _RET_IP_);
 233         return state;
 234 }
 235
 236 void free_extent_state(struct extent_state *state)
 237 {
 238         if (!state)
 239                 return;
 240         if (refcount_dec_and_test(&state->refs)) {
 241                 WARN_ON(extent_state_in_tree(state));
 242                 btrfs_leak_debug_del(&state->leak_list);
 243                 trace_free_extent_state(state, _RET_IP_);
 244                 kmem_cache_free(extent_state_cache, state);
 245         }
 246 }
 247
 248 static struct rb_node *tree_insert(struct rb_root *root,
 249                                    struct rb_node *search_start,
 250                                    u64 offset,
 251                                    struct rb_node *node,
 252                                    struct rb_node ***p_in,
 253                                    struct rb_node **parent_in)
 254 {
 255         struct rb_node **p;
 256         struct rb_node *parent = NULL;
 257         struct tree_entry *entry;
 258
 259         if (p_in && parent_in) {
 260                 p = *p_in;
 261                 parent = *parent_in;
 262                 goto do_insert;
 263         }
 264
 265         p = search_start ? &search_start : &root->rb_node;
 266         while (*p) {
 267                 parent = *p;
 268                 entry = rb_entry(parent, struct tree_entry, rb_node);
 269
 270                 if (offset < entry->start)
 271                         p = &(*p)->rb_left;
 272                 else if (offset > entry->end)
 273                         p = &(*p)->rb_right;
 274                 else
 275                         return parent;
 276         }
 277
 278 do_insert:
 279         rb_link_node(node, parent, p);
 280         rb_insert_color(node, root);
 281         return NULL;
 282 }
 283
 284 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 285                                       struct rb_node **prev_ret,
 286                                       struct rb_node **next_ret,
 287                                       struct rb_node ***p_ret,
 288                                       struct rb_node **parent_ret)
 289 {
 290         struct rb_root *root = &tree->state;
 291         struct rb_node **n = &root->rb_node;
 292         struct rb_node *prev = NULL;
 293         struct rb_node *orig_prev = NULL;
 294         struct tree_entry *entry;
 295         struct tree_entry *prev_entry = NULL;
 296
 297         while (*n) {
 298                 prev = *n;
 299                 entry = rb_entry(prev, struct tree_entry, rb_node);
 300                 prev_entry = entry;
 301
 302                 if (offset < entry->start)
 303                         n = &(*n)->rb_left;
 304                 else if (offset > entry->end)
 305                         n = &(*n)->rb_right;
 306                 else
 307                         return *n;
 308         }
 309
 310         if (p_ret)
 311                 *p_ret = n;
 312         if (parent_ret)
 313                 *parent_ret = prev;
 314
 315         if (prev_ret) {
 316                 orig_prev = prev;
 317                 while (prev && offset > prev_entry->end) {
 318                         prev = rb_next(prev);
 319                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 320                 }
 321                 *prev_ret = prev;
 322                 prev = orig_prev;
 323         }
 324
 325         if (next_ret) {
 326                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 327                 while (prev && offset < prev_entry->start) {
 328                         prev = rb_prev(prev);
 329                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 330                 }
 331                 *next_ret = prev;
 332         }
 333         return NULL;
 334 }
 335
 336 static inline struct rb_node *
 337 tree_search_for_insert(struct extent_io_tree *tree,
 338                        u64 offset,
 339                        struct rb_node ***p_ret,
 340                        struct rb_node **parent_ret)
 341 {
 342         struct rb_node *prev = NULL;
 343         struct rb_node *ret;
 344
 345         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 346         if (!ret)
 347                 return prev;
 348         return ret;
 349 }
 350
 351 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 352                                           u64 offset)
 353 {
 354         return tree_search_for_insert(tree, offset, NULL, NULL);
 355 }
 356
 357 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 358                      struct extent_state *other)
 359 {
 360         if (tree->ops && tree->ops->merge_extent_hook)
 361                 tree->ops->merge_extent_hook(tree->private_data, new, other);
 362 }
 363
 364 /*
 365  * utility function to look for merge candidates inside a given range.
 366  * Any extents with matching state are merged together into a single
 367  * extent in the tree.  Extents with EXTENT_IO in their state field
 368  * are not merged because the end_io handlers need to be able to do
 369  * operations on them without sleeping (or doing allocations/splits).
 370  *
 371  * This should be called with the tree lock held.
 372  */
 373 static void merge_state(struct extent_io_tree *tree,
 374                         struct extent_state *state)
 375 {
 376         struct extent_state *other;
 377         struct rb_node *other_node;
 378
 379         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 380                 return;
 381
 382         other_node = rb_prev(&state->rb_node);
 383         if (other_node) {
 384                 other = rb_entry(other_node, struct extent_state, rb_node);
 385                 if (other->end == state->start - 1 &&
 386                     other->state == state->state) {
 387                         merge_cb(tree, state, other);
 388                         state->start = other->start;
 389                         rb_erase(&other->rb_node, &tree->state);
 390                         RB_CLEAR_NODE(&other->rb_node);
 391                         free_extent_state(other);
 392                 }
 393         }
 394         other_node = rb_next(&state->rb_node);
 395         if (other_node) {
 396                 other = rb_entry(other_node, struct extent_state, rb_node);
 397                 if (other->start == state->end + 1 &&
 398                     other->state == state->state) {
 399                         merge_cb(tree, state, other);
 400                         state->end = other->end;
 401                         rb_erase(&other->rb_node, &tree->state);
 402                         RB_CLEAR_NODE(&other->rb_node);
 403                         free_extent_state(other);
 404                 }
 405         }
 406 }
 407
 408 static void set_state_cb(struct extent_io_tree *tree,
 409                          struct extent_state *state, unsigned *bits)
 410 {
 411         if (tree->ops && tree->ops->set_bit_hook)
 412                 tree->ops->set_bit_hook(tree->private_data, state, bits);
 413 }
 414
 415 static void clear_state_cb(struct extent_io_tree *tree,
 416                            struct extent_state *state, unsigned *bits)
 417 {
 418         if (tree->ops && tree->ops->clear_bit_hook)
 419                 tree->ops->clear_bit_hook(tree->private_data, state, bits);
 420 }
 421
 422 static void set_state_bits(struct extent_io_tree *tree,
 423                            struct extent_state *state, unsigned *bits,
 424                            struct extent_changeset *changeset);
 425
 426 /*
 427  * insert an extent_state struct into the tree.  'bits' are set on the
 428  * struct before it is inserted.
 429  *
 430  * This may return -EEXIST if the extent is already there, in which case the
 431  * state struct is freed.
 432  *
 433  * The tree lock is not taken internally.  This is a utility function and
 434  * probably isn't what you want to call (see set/clear_extent_bit).
 435  */
 436 static int insert_state(struct extent_io_tree *tree,
 437                         struct extent_state *state, u64 start, u64 end,
 438                         struct rb_node ***p,
 439                         struct rb_node **parent,
 440                         unsigned *bits, struct extent_changeset *changeset)
 441 {
 442         struct rb_node *node;
 443
 444         if (end < start)
 445                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 446                        end, start);
 447         state->start = start;
 448         state->end = end;
 449
 450         set_state_bits(tree, state, bits, changeset);
 451
 452         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 453         if (node) {
 454                 struct extent_state *found;
 455                 found = rb_entry(node, struct extent_state, rb_node);
 456                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
 457                        found->start, found->end, start, end);
 458                 return -EEXIST;
 459         }
 460         merge_state(tree, state);
 461         return 0;
 462 }
 463
 464 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 465                      u64 split)
 466 {
 467         if (tree->ops && tree->ops->split_extent_hook)
 468                 tree->ops->split_extent_hook(tree->private_data, orig, split);
 469 }
 470
 471 /*
 472  * split a given extent state struct in two, inserting the preallocated
 473  * struct 'prealloc' as the newly created second half.  'split' indicates an
 474  * offset inside 'orig' where it should be split.
 475  *
 476  * Before calling,
 477  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 478  * are two extent state structs in the tree:
 479  * prealloc: [orig->start, split - 1]
 480  * orig: [ split, orig->end ]
 481  *
 482  * The tree locks are not taken by this function. They need to be held
 483  * by the caller.
 484  */
 485 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 486                        struct extent_state *prealloc, u64 split)
 487 {
 488         struct rb_node *node;
 489
 490         split_cb(tree, orig, split);
 491
 492         prealloc->start = orig->start;
 493         prealloc->end = split - 1;
 494         prealloc->state = orig->state;
 495         orig->start = split;
 496
 497         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 498                            &prealloc->rb_node, NULL, NULL);
 499         if (node) {
 500                 free_extent_state(prealloc);
 501                 return -EEXIST;
 502         }
 503         return 0;
 504 }
 505
 506 static struct extent_state *next_state(struct extent_state *state)
 507 {
 508         struct rb_node *next = rb_next(&state->rb_node);
 509         if (next)
 510                 return rb_entry(next, struct extent_state, rb_node);
 511         else
 512                 return NULL;
 513 }
 514
 515 /*
 516  * utility function to clear some bits in an extent state struct.
 517  * it will optionally wake up any one waiting on this state (wake == 1).
 518  *
 519  * If no bits are set on the state struct after clearing things, the
 520  * struct is freed and removed from the tree
 521  */
 522 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 523                                             struct extent_state *state,
 524                                             unsigned *bits, int wake,
 525                                             struct extent_changeset *changeset)
 526 {
 527         struct extent_state *next;
 528         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 529
 530         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 531                 u64 range = state->end - state->start + 1;
 532                 WARN_ON(range > tree->dirty_bytes);
 533                 tree->dirty_bytes -= range;
 534         }
 535         clear_state_cb(tree, state, bits);
 536         add_extent_changeset(state, bits_to_clear, changeset, 0);
 537         state->state &= ~bits_to_clear;
 538         if (wake)
 539                 wake_up(&state->wq);
 540         if (state->state == 0) {
 541                 next = next_state(state);
 542                 if (extent_state_in_tree(state)) {
 543                         rb_erase(&state->rb_node, &tree->state);
 544                         RB_CLEAR_NODE(&state->rb_node);
 545                         free_extent_state(state);
 546                 } else {
 547                         WARN_ON(1);
 548                 }
 549         } else {
 550                 merge_state(tree, state);
 551                 next = next_state(state);
 552         }
 553         return next;
 554 }
 555
 556 static struct extent_state *
 557 alloc_extent_state_atomic(struct extent_state *prealloc)
 558 {
 559         if (!prealloc)
 560                 prealloc = alloc_extent_state(GFP_ATOMIC);
 561
 562         return prealloc;
 563 }
 564
 565 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 566 {
 567         btrfs_panic(tree_fs_info(tree), err,
 568                     "Locking error: Extent tree was modified by another thread while locked.");
 569 }
 570
 571 /*
 572  * clear some bits on a range in the tree.  This may require splitting
 573  * or inserting elements in the tree, so the gfp mask is used to
 574  * indicate which allocations or sleeping are allowed.
 575  *
 576  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 577  * the given range from the tree regardless of state (ie for truncate).
 578  *
 579  * the range [start, end] is inclusive.
 580  *
 581  * This takes the tree lock, and returns 0 on success and < 0 on error.
 582  */
 583 int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 584                               unsigned bits, int wake, int delete,
 585                               struct extent_state **cached_state,
 586                               gfp_t mask, struct extent_changeset *changeset)
 587 {
 588         struct extent_state *state;
 589         struct extent_state *cached;
 590         struct extent_state *prealloc = NULL;
 591         struct rb_node *node;
 592         u64 last_end;
 593         int err;
 594         int clear = 0;
 595
 596         btrfs_debug_check_extent_io_range(tree, start, end);
 597
 598         if (bits & EXTENT_DELALLOC)
 599                 bits |= EXTENT_NORESERVE;
 600
 601         if (delete)
 602                 bits |= ~EXTENT_CTLBITS;
 603         bits |= EXTENT_FIRST_DELALLOC;
 604
 605         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 606                 clear = 1;
 607 again:
 608         if (!prealloc && gfpflags_allow_blocking(mask)) {
 609                 /*
 610                  * Don't care for allocation failure here because we might end
 611                  * up not needing the pre-allocated extent state at all, which
 612                  * is the case if we only have in the tree extent states that
 613                  * cover our input range and don't cover too any other range.
 614                  * If we end up needing a new extent state we allocate it later.
 615                  */
 616                 prealloc = alloc_extent_state(mask);
 617         }
 618
 619         spin_lock(&tree->lock);
 620         if (cached_state) {
 621                 cached = *cached_state;
 622
 623                 if (clear) {
 624                         *cached_state = NULL;
 625                         cached_state = NULL;
 626                 }
 627
 628                 if (cached && extent_state_in_tree(cached) &&
 629                     cached->start <= start && cached->end > start) {
 630                         if (clear)
 631                                 refcount_dec(&cached->refs);
 632                         state = cached;
 633                         goto hit_next;
 634                 }
 635                 if (clear)
 636                         free_extent_state(cached);
 637         }
 638         /*
 639          * this search will find the extents that end after
 640          * our range starts
 641          */
 642         node = tree_search(tree, start);
 643         if (!node)
 644                 goto out;
 645         state = rb_entry(node, struct extent_state, rb_node);
 646 hit_next:
 647         if (state->start > end)
 648                 goto out;
 649         WARN_ON(state->end < start);
 650         last_end = state->end;
 651
 652         /* the state doesn't have the wanted bits, go ahead */
 653         if (!(state->state & bits)) {
 654                 state = next_state(state);
 655                 goto next;
 656         }
 657
 658         /*
 659          *     | ---- desired range ---- |
 660          *  | state | or
 661          *  | ------------- state -------------- |
 662          *
 663          * We need to split the extent we found, and may flip
 664          * bits on second half.
 665          *
 666          * If the extent we found extends past our range, we
 667          * just split and search again.  It'll get split again
 668          * the next time though.
 669          *
 670          * If the extent we found is inside our range, we clear
 671          * the desired bit on it.
 672          */
 673
 674         if (state->start < start) {
 675                 prealloc = alloc_extent_state_atomic(prealloc);
 676                 BUG_ON(!prealloc);
 677                 err = split_state(tree, state, prealloc, start);
 678                 if (err)
 679                         extent_io_tree_panic(tree, err);
 680
 681                 prealloc = NULL;
 682                 if (err)
 683                         goto out;
 684                 if (state->end <= end) {
 685                         state = clear_state_bit(tree, state, &bits, wake,
 686                                                 changeset);
 687                         goto next;
 688                 }
 689                 goto search_again;
 690         }
 691         /*
 692          * | ---- desired range ---- |
 693          *                        | state |
 694          * We need to split the extent, and clear the bit
 695          * on the first half
 696          */
 697         if (state->start <= end && state->end > end) {
 698                 prealloc = alloc_extent_state_atomic(prealloc);
 699                 BUG_ON(!prealloc);
 700                 err = split_state(tree, state, prealloc, end + 1);
 701                 if (err)
 702                         extent_io_tree_panic(tree, err);
 703
 704                 if (wake)
 705                         wake_up(&state->wq);
 706
 707                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 708
 709                 prealloc = NULL;
 710                 goto out;
 711         }
 712
 713         state = clear_state_bit(tree, state, &bits, wake, changeset);
 714 next:
 715         if (last_end == (u64)-1)
 716                 goto out;
 717         start = last_end + 1;
 718         if (start <= end && state && !need_resched())
 719                 goto hit_next;
 720
 721 search_again:
 722         if (start > end)
 723                 goto out;
 724         spin_unlock(&tree->lock);
 725         if (gfpflags_allow_blocking(mask))
 726                 cond_resched();
 727         goto again;
 728
 729 out:
 730         spin_unlock(&tree->lock);
 731         if (prealloc)
 732                 free_extent_state(prealloc);
 733
 734         return 0;
 735
 736 }
 737
 738 static void wait_on_state(struct extent_io_tree *tree,
 739                           struct extent_state *state)
 740                 __releases(tree->lock)
 741                 __acquires(tree->lock)
 742 {
 743         DEFINE_WAIT(wait);
 744         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 745         spin_unlock(&tree->lock);
 746         schedule();
 747         spin_lock(&tree->lock);
 748         finish_wait(&state->wq, &wait);
 749 }
 750
 751 /*
 752  * waits for one or more bits to clear on a range in the state tree.
 753  * The range [start, end] is inclusive.
 754  * The tree lock is taken by this function
 755  */
 756 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 757                             unsigned long bits)
 758 {
 759         struct extent_state *state;
 760         struct rb_node *node;
 761
 762         btrfs_debug_check_extent_io_range(tree, start, end);
 763
 764         spin_lock(&tree->lock);
 765 again:
 766         while (1) {
 767                 /*
 768                  * this search will find all the extents that end after
 769                  * our range starts
 770                  */
 771                 node = tree_search(tree, start);
 772 process_node:
 773                 if (!node)
 774                         break;
 775
 776                 state = rb_entry(node, struct extent_state, rb_node);
 777
 778                 if (state->start > end)
 779                         goto out;
 780
 781                 if (state->state & bits) {
 782                         start = state->start;
 783                         refcount_inc(&state->refs);
 784                         wait_on_state(tree, state);
 785                         free_extent_state(state);
 786                         goto again;
 787                 }
 788                 start = state->end + 1;
 789
 790                 if (start > end)
 791                         break;
 792
 793                 if (!cond_resched_lock(&tree->lock)) {
 794                         node = rb_next(node);
 795                         goto process_node;
 796                 }
 797         }
 798 out:
 799         spin_unlock(&tree->lock);
 800 }
 801
 802 static void set_state_bits(struct extent_io_tree *tree,
 803                            struct extent_state *state,
 804                            unsigned *bits, struct extent_changeset *changeset)
 805 {
 806         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 807
 808         set_state_cb(tree, state, bits);
 809         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 810                 u64 range = state->end - state->start + 1;
 811                 tree->dirty_bytes += range;
 812         }
 813         add_extent_changeset(state, bits_to_set, changeset, 1);
 814         state->state |= bits_to_set;
 815 }
 816
 817 static void cache_state_if_flags(struct extent_state *state,
 818                                  struct extent_state **cached_ptr,
 819                                  unsigned flags)
 820 {
 821         if (cached_ptr && !(*cached_ptr)) {
 822                 if (!flags || (state->state & flags)) {
 823                         *cached_ptr = state;
 824                         refcount_inc(&state->refs);
 825                 }
 826         }
 827 }
 828
 829 static void cache_state(struct extent_state *state,
 830                         struct extent_state **cached_ptr)
 831 {
 832         return cache_state_if_flags(state, cached_ptr,
 833                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 834 }
 835
 836 /*
 837  * set some bits on a range in the tree.  This may require allocations or
 838  * sleeping, so the gfp mask is used to indicate what is allowed.
 839  *
 840  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 841  * part of the range already has the desired bits set.  The start of the
 842  * existing range is returned in failed_start in this case.
 843  *
 844  * [start, end] is inclusive This takes the tree lock.
 845  */
 846
 847 static int __must_check
 848 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 849                  unsigned bits, unsigned exclusive_bits,
 850                  u64 *failed_start, struct extent_state **cached_state,
 851                  gfp_t mask, struct extent_changeset *changeset)
 852 {
 853         struct extent_state *state;
 854         struct extent_state *prealloc = NULL;
 855         struct rb_node *node;
 856         struct rb_node **p;
 857         struct rb_node *parent;
 858         int err = 0;
 859         u64 last_start;
 860         u64 last_end;
 861
 862         btrfs_debug_check_extent_io_range(tree, start, end);
 863
 864         bits |= EXTENT_FIRST_DELALLOC;
 865 again:
 866         if (!prealloc && gfpflags_allow_blocking(mask)) {
 867                 /*
 868                  * Don't care for allocation failure here because we might end
 869                  * up not needing the pre-allocated extent state at all, which
 870                  * is the case if we only have in the tree extent states that
 871                  * cover our input range and don't cover too any other range.
 872                  * If we end up needing a new extent state we allocate it later.
 873                  */
 874                 prealloc = alloc_extent_state(mask);
 875         }
 876
 877         spin_lock(&tree->lock);
 878         if (cached_state && *cached_state) {
 879                 state = *cached_state;
 880                 if (state->start <= start && state->end > start &&
 881                     extent_state_in_tree(state)) {
 882                         node = &state->rb_node;
 883                         goto hit_next;
 884                 }
 885         }
 886         /*
 887          * this search will find all the extents that end after
 888          * our range starts.
 889          */
 890         node = tree_search_for_insert(tree, start, &p, &parent);
 891         if (!node) {
 892                 prealloc = alloc_extent_state_atomic(prealloc);
 893                 BUG_ON(!prealloc);
 894                 err = insert_state(tree, prealloc, start, end,
 895                                    &p, &parent, &bits, changeset);
 896                 if (err)
 897                         extent_io_tree_panic(tree, err);
 898
 899                 cache_state(prealloc, cached_state);
 900                 prealloc = NULL;
 901                 goto out;
 902         }
 903         state = rb_entry(node, struct extent_state, rb_node);
 904 hit_next:
 905         last_start = state->start;
 906         last_end = state->end;
 907
 908         /*
 909          * | ---- desired range ---- |
 910          * | state |
 911          *
 912          * Just lock what we found and keep going
 913          */
 914         if (state->start == start && state->end <= end) {
 915                 if (state->state & exclusive_bits) {
 916                         *failed_start = state->start;
 917                         err = -EEXIST;
 918                         goto out;
 919                 }
 920
 921                 set_state_bits(tree, state, &bits, changeset);
 922                 cache_state(state, cached_state);
 923                 merge_state(tree, state);
 924                 if (last_end == (u64)-1)
 925                         goto out;
 926                 start = last_end + 1;
 927                 state = next_state(state);
 928                 if (start < end && state && state->start == start &&
 929                     !need_resched())
 930                         goto hit_next;
 931                 goto search_again;
 932         }
 933
 934         /*
 935          *     | ---- desired range ---- |
 936          * | state |
 937          *   or
 938          * | ------------- state -------------- |
 939          *
 940          * We need to split the extent we found, and may flip bits on
 941          * second half.
 942          *
 943          * If the extent we found extends past our
 944          * range, we just split and search again.  It'll get split
 945          * again the next time though.
 946          *
 947          * If the extent we found is inside our range, we set the
 948          * desired bit on it.
 949          */
 950         if (state->start < start) {
 951                 if (state->state & exclusive_bits) {
 952                         *failed_start = start;
 953                         err = -EEXIST;
 954                         goto out;
 955                 }
 956
 957                 prealloc = alloc_extent_state_atomic(prealloc);
 958                 BUG_ON(!prealloc);
 959                 err = split_state(tree, state, prealloc, start);
 960                 if (err)
 961                         extent_io_tree_panic(tree, err);
 962
 963                 prealloc = NULL;
 964                 if (err)
 965                         goto out;
 966                 if (state->end <= end) {
 967                         set_state_bits(tree, state, &bits, changeset);
 968                         cache_state(state, cached_state);
 969                         merge_state(tree, state);
 970                         if (last_end == (u64)-1)
 971                                 goto out;
 972                         start = last_end + 1;
 973                         state = next_state(state);
 974                         if (start < end && state && state->start == start &&
 975                             !need_resched())
 976                                 goto hit_next;
 977                 }
 978                 goto search_again;
 979         }
 980         /*
 981          * | ---- desired range ---- |
 982          *     | state | or               | state |
 983          *
 984          * There's a hole, we need to insert something in it and
 985          * ignore the extent we found.
 986          */
 987         if (state->start > start) {
 988                 u64 this_end;
 989                 if (end < last_start)
 990                         this_end = end;
 991                 else
 992                         this_end = last_start - 1;
 993
 994                 prealloc = alloc_extent_state_atomic(prealloc);
 995                 BUG_ON(!prealloc);
 996
 997                 /*
 998                  * Avoid to free 'prealloc' if it can be merged with
 999                  * the later extent.
1000                  */
1001                 err = insert_state(tree, prealloc, start, this_end,
1002                                    NULL, NULL, &bits, changeset);
1003                 if (err)
1004                         extent_io_tree_panic(tree, err);
1005
1006                 cache_state(prealloc, cached_state);
1007                 prealloc = NULL;
1008                 start = this_end + 1;
1009                 goto search_again;
1010         }
1011         /*
1012          * | ---- desired range ---- |
1013          *                        | state |
1014          * We need to split the extent, and set the bit
1015          * on the first half
1016          */
1017         if (state->start <= end && state->end > end) {
1018                 if (state->state & exclusive_bits) {
1019                         *failed_start = start;
1020                         err = -EEXIST;
1021                         goto out;
1022                 }
1023
1024                 prealloc = alloc_extent_state_atomic(prealloc);
1025                 BUG_ON(!prealloc);
1026                 err = split_state(tree, state, prealloc, end + 1);
1027                 if (err)
1028                         extent_io_tree_panic(tree, err);
1029
1030                 set_state_bits(tree, prealloc, &bits, changeset);
1031                 cache_state(prealloc, cached_state);
1032                 merge_state(tree, prealloc);
1033                 prealloc = NULL;
1034                 goto out;
1035         }
1036
1037 search_again:
1038         if (start > end)
1039                 goto out;
1040         spin_unlock(&tree->lock);
1041         if (gfpflags_allow_blocking(mask))
1042                 cond_resched();
1043         goto again;
1044
1045 out:
1046         spin_unlock(&tree->lock);
1047         if (prealloc)
1048                 free_extent_state(prealloc);
1049
1050         return err;
1051
1052 }
1053
1054 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1055                    unsigned bits, u64 * failed_start,
1056                    struct extent_state **cached_state, gfp_t mask)
1057 {
1058         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1059                                 cached_state, mask, NULL);
1060 }
1061
1062
1063 /**
1064  * convert_extent_bit - convert all bits in a given range from one bit to
1065  *                      another
1066  * @tree:       the io tree to search
1067  * @start:      the start offset in bytes
1068  * @end:        the end offset in bytes (inclusive)
1069  * @bits:       the bits to set in this range
1070  * @clear_bits: the bits to clear in this range
1071  * @cached_state:       state that we're going to cache
1072  *
1073  * This will go through and set bits for the given range.  If any states exist
1074  * already in this range they are set with the given bit and cleared of the
1075  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1076  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1077  * boundary bits like LOCK.
1078  *
1079  * All allocations are done with GFP_NOFS.
1080  */
1081 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1082                        unsigned bits, unsigned clear_bits,
1083                        struct extent_state **cached_state)
1084 {
1085         struct extent_state *state;
1086         struct extent_state *prealloc = NULL;
1087         struct rb_node *node;
1088         struct rb_node **p;
1089         struct rb_node *parent;
1090         int err = 0;
1091         u64 last_start;
1092         u64 last_end;
1093         bool first_iteration = true;
1094
1095         btrfs_debug_check_extent_io_range(tree, start, end);
1096
1097 again:
1098         if (!prealloc) {
1099                 /*
1100                  * Best effort, don't worry if extent state allocation fails
1101                  * here for the first iteration. We might have a cached state
1102                  * that matches exactly the target range, in which case no
1103                  * extent state allocations are needed. We'll only know this
1104                  * after locking the tree.
1105                  */
1106                 prealloc = alloc_extent_state(GFP_NOFS);
1107                 if (!prealloc && !first_iteration)
1108                         return -ENOMEM;
1109         }
1110
1111         spin_lock(&tree->lock);
1112         if (cached_state && *cached_state) {
1113                 state = *cached_state;
1114                 if (state->start <= start && state->end > start &&
1115                     extent_state_in_tree(state)) {
1116                         node = &state->rb_node;
1117                         goto hit_next;
1118                 }
1119         }
1120
1121         /*
1122          * this search will find all the extents that end after
1123          * our range starts.
1124          */
1125         node = tree_search_for_insert(tree, start, &p, &parent);
1126         if (!node) {
1127                 prealloc = alloc_extent_state_atomic(prealloc);
1128                 if (!prealloc) {
1129                         err = -ENOMEM;
1130                         goto out;
1131                 }
1132                 err = insert_state(tree, prealloc, start, end,
1133                                    &p, &parent, &bits, NULL);
1134                 if (err)
1135                         extent_io_tree_panic(tree, err);
1136                 cache_state(prealloc, cached_state);
1137                 prealloc = NULL;
1138                 goto out;
1139         }
1140         state = rb_entry(node, struct extent_state, rb_node);
1141 hit_next:
1142         last_start = state->start;
1143         last_end = state->end;
1144
1145         /*
1146          * | ---- desired range ---- |
1147          * | state |
1148          *
1149          * Just lock what we found and keep going
1150          */
1151         if (state->start == start && state->end <= end) {
1152                 set_state_bits(tree, state, &bits, NULL);
1153                 cache_state(state, cached_state);
1154                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1155                 if (last_end == (u64)-1)
1156                         goto out;
1157                 start = last_end + 1;
1158                 if (start < end && state && state->start == start &&
1159                     !need_resched())
1160                         goto hit_next;
1161                 goto search_again;
1162         }
1163
1164         /*
1165          *     | ---- desired range ---- |
1166          * | state |
1167          *   or
1168          * | ------------- state -------------- |
1169          *
1170          * We need to split the extent we found, and may flip bits on
1171          * second half.
1172          *
1173          * If the extent we found extends past our
1174          * range, we just split and search again.  It'll get split
1175          * again the next time though.
1176          *
1177          * If the extent we found is inside our range, we set the
1178          * desired bit on it.
1179          */
1180         if (state->start < start) {
1181                 prealloc = alloc_extent_state_atomic(prealloc);
1182                 if (!prealloc) {
1183                         err = -ENOMEM;
1184                         goto out;
1185                 }
1186                 err = split_state(tree, state, prealloc, start);
1187                 if (err)
1188                         extent_io_tree_panic(tree, err);
1189                 prealloc = NULL;
1190                 if (err)
1191                         goto out;
1192                 if (state->end <= end) {
1193                         set_state_bits(tree, state, &bits, NULL);
1194                         cache_state(state, cached_state);
1195                         state = clear_state_bit(tree, state, &clear_bits, 0,
1196                                                 NULL);
1197                         if (last_end == (u64)-1)
1198                                 goto out;
1199                         start = last_end + 1;
1200                         if (start < end && state && state->start == start &&
1201                             !need_resched())
1202                                 goto hit_next;
1203                 }
1204                 goto search_again;
1205         }
1206         /*
1207          * | ---- desired range ---- |
1208          *     | state | or               | state |
1209          *
1210          * There's a hole, we need to insert something in it and
1211          * ignore the extent we found.
1212          */
1213         if (state->start > start) {
1214                 u64 this_end;
1215                 if (end < last_start)
1216                         this_end = end;
1217                 else
1218                         this_end = last_start - 1;
1219
1220                 prealloc = alloc_extent_state_atomic(prealloc);
1221                 if (!prealloc) {
1222                         err = -ENOMEM;
1223                         goto out;
1224                 }
1225
1226                 /*
1227                  * Avoid to free 'prealloc' if it can be merged with
1228                  * the later extent.
1229                  */
1230                 err = insert_state(tree, prealloc, start, this_end,
1231                                    NULL, NULL, &bits, NULL);
1232                 if (err)
1233                         extent_io_tree_panic(tree, err);
1234                 cache_state(prealloc, cached_state);
1235                 prealloc = NULL;
1236                 start = this_end + 1;
1237                 goto search_again;
1238         }
1239         /*
1240          * | ---- desired range ---- |
1241          *                        | state |
1242          * We need to split the extent, and set the bit
1243          * on the first half
1244          */
1245         if (state->start <= end && state->end > end) {
1246                 prealloc = alloc_extent_state_atomic(prealloc);
1247                 if (!prealloc) {
1248                         err = -ENOMEM;
1249                         goto out;
1250                 }
1251
1252                 err = split_state(tree, state, prealloc, end + 1);
1253                 if (err)
1254                         extent_io_tree_panic(tree, err);
1255
1256                 set_state_bits(tree, prealloc, &bits, NULL);
1257                 cache_state(prealloc, cached_state);
1258                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1259                 prealloc = NULL;
1260                 goto out;
1261         }
1262
1263 search_again:
1264         if (start > end)
1265                 goto out;
1266         spin_unlock(&tree->lock);
1267         cond_resched();
1268         first_iteration = false;
1269         goto again;
1270
1271 out:
1272         spin_unlock(&tree->lock);
1273         if (prealloc)
1274                 free_extent_state(prealloc);
1275
1276         return err;
1277 }
1278
1279 /* wrappers around set/clear extent bit */
1280 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1281                            unsigned bits, struct extent_changeset *changeset)
1282 {
1283         /*
1284          * We don't support EXTENT_LOCKED yet, as current changeset will
1285          * record any bits changed, so for EXTENT_LOCKED case, it will
1286          * either fail with -EEXIST or changeset will record the whole
1287          * range.
1288          */
1289         BUG_ON(bits & EXTENT_LOCKED);
1290
1291         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1292                                 changeset);
1293 }
1294
1295 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1296                      unsigned bits, int wake, int delete,
1297                      struct extent_state **cached)
1298 {
1299         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1300                                   cached, GFP_NOFS, NULL);
1301 }
1302
1303 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1304                 unsigned bits, struct extent_changeset *changeset)
1305 {
1306         /*
1307          * Don't support EXTENT_LOCKED case, same reason as
1308          * set_record_extent_bits().
1309          */
1310         BUG_ON(bits & EXTENT_LOCKED);
1311
1312         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1313                                   changeset);
1314 }
1315
1316 /*
1317  * either insert or lock state struct between start and end use mask to tell
1318  * us if waiting is desired.
1319  */
1320 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1321                      struct extent_state **cached_state)
1322 {
1323         int err;
1324         u64 failed_start;
1325
1326         while (1) {
1327                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1328                                        EXTENT_LOCKED, &failed_start,
1329                                        cached_state, GFP_NOFS, NULL);
1330                 if (err == -EEXIST) {
1331                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1332                         start = failed_start;
1333                 } else
1334                         break;
1335                 WARN_ON(start > end);
1336         }
1337         return err;
1338 }
1339
1340 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1341 {
1342         int err;
1343         u64 failed_start;
1344
1345         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1346                                &failed_start, NULL, GFP_NOFS, NULL);
1347         if (err == -EEXIST) {
1348                 if (failed_start > start)
1349                         clear_extent_bit(tree, start, failed_start - 1,
1350                                          EXTENT_LOCKED, 1, 0, NULL);
1351                 return 0;
1352         }
1353         return 1;
1354 }
1355
1356 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1357 {
1358         unsigned long index = start >> PAGE_SHIFT;
1359         unsigned long end_index = end >> PAGE_SHIFT;
1360         struct page *page;
1361
1362         while (index <= end_index) {
1363                 page = find_get_page(inode->i_mapping, index);
1364                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1365                 clear_page_dirty_for_io(page);
1366                 put_page(page);
1367                 index++;
1368         }
1369 }
1370
1371 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1372 {
1373         unsigned long index = start >> PAGE_SHIFT;
1374         unsigned long end_index = end >> PAGE_SHIFT;
1375         struct page *page;
1376
1377         while (index <= end_index) {
1378                 page = find_get_page(inode->i_mapping, index);
1379                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1380                 __set_page_dirty_nobuffers(page);
1381                 account_page_redirty(page);
1382                 put_page(page);
1383                 index++;
1384         }
1385 }
1386
1387 /*
1388  * helper function to set both pages and extents in the tree writeback
1389  */
1390 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1391 {
1392         tree->ops->set_range_writeback(tree->private_data, start, end);
1393 }
1394
1395 /* find the first state struct with 'bits' set after 'start', and
1396  * return it.  tree->lock must be held.  NULL will returned if
1397  * nothing was found after 'start'
1398  */
1399 static struct extent_state *
1400 find_first_extent_bit_state(struct extent_io_tree *tree,
1401                             u64 start, unsigned bits)
1402 {
1403         struct rb_node *node;
1404         struct extent_state *state;
1405
1406         /*
1407          * this search will find all the extents that end after
1408          * our range starts.
1409          */
1410         node = tree_search(tree, start);
1411         if (!node)
1412                 goto out;
1413
1414         while (1) {
1415                 state = rb_entry(node, struct extent_state, rb_node);
1416                 if (state->end >= start && (state->state & bits))
1417                         return state;
1418
1419                 node = rb_next(node);
1420                 if (!node)
1421                         break;
1422         }
1423 out:
1424         return NULL;
1425 }
1426
1427 /*
1428  * find the first offset in the io tree with 'bits' set. zero is
1429  * returned if we find something, and *start_ret and *end_ret are
1430  * set to reflect the state struct that was found.
1431  *
1432  * If nothing was found, 1 is returned. If found something, return 0.
1433  */
1434 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1435                           u64 *start_ret, u64 *end_ret, unsigned bits,
1436                           struct extent_state **cached_state)
1437 {
1438         struct extent_state *state;
1439         struct rb_node *n;
1440         int ret = 1;
1441
1442         spin_lock(&tree->lock);
1443         if (cached_state && *cached_state) {
1444                 state = *cached_state;
1445                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1446                         n = rb_next(&state->rb_node);
1447                         while (n) {
1448                                 state = rb_entry(n, struct extent_state,
1449                                                  rb_node);
1450                                 if (state->state & bits)
1451                                         goto got_it;
1452                                 n = rb_next(n);
1453                         }
1454                         free_extent_state(*cached_state);
1455                         *cached_state = NULL;
1456                         goto out;
1457                 }
1458                 free_extent_state(*cached_state);
1459                 *cached_state = NULL;
1460         }
1461
1462         state = find_first_extent_bit_state(tree, start, bits);
1463 got_it:
1464         if (state) {
1465                 cache_state_if_flags(state, cached_state, 0);
1466                 *start_ret = state->start;
1467                 *end_ret = state->end;
1468                 ret = 0;
1469         }
1470 out:
1471         spin_unlock(&tree->lock);
1472         return ret;
1473 }
1474
1475 /*
1476  * find a contiguous range of bytes in the file marked as delalloc, not
1477  * more than 'max_bytes'.  start and end are used to return the range,
1478  *
1479  * 1 is returned if we find something, 0 if nothing was in the tree
1480  */
1481 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1482                                         u64 *start, u64 *end, u64 max_bytes,
1483                                         struct extent_state **cached_state)
1484 {
1485         struct rb_node *node;
1486         struct extent_state *state;
1487         u64 cur_start = *start;
1488         u64 found = 0;
1489         u64 total_bytes = 0;
1490
1491         spin_lock(&tree->lock);
1492
1493         /*
1494          * this search will find all the extents that end after
1495          * our range starts.
1496          */
1497         node = tree_search(tree, cur_start);
1498         if (!node) {
1499                 if (!found)
1500                         *end = (u64)-1;
1501                 goto out;
1502         }
1503
1504         while (1) {
1505                 state = rb_entry(node, struct extent_state, rb_node);
1506                 if (found && (state->start != cur_start ||
1507                               (state->state & EXTENT_BOUNDARY))) {
1508                         goto out;
1509                 }
1510                 if (!(state->state & EXTENT_DELALLOC)) {
1511                         if (!found)
1512                                 *end = state->end;
1513                         goto out;
1514                 }
1515                 if (!found) {
1516                         *start = state->start;
1517                         *cached_state = state;
1518                         refcount_inc(&state->refs);
1519                 }
1520                 found++;
1521                 *end = state->end;
1522                 cur_start = state->end + 1;
1523                 node = rb_next(node);
1524                 total_bytes += state->end - state->start + 1;
1525                 if (total_bytes >= max_bytes)
1526                         break;
1527                 if (!node)
1528                         break;
1529         }
1530 out:
1531         spin_unlock(&tree->lock);
1532         return found;
1533 }
1534
1535 static int __process_pages_contig(struct address_space *mapping,
1536                                   struct page *locked_page,
1537                                   pgoff_t start_index, pgoff_t end_index,
1538                                   unsigned long page_ops, pgoff_t *index_ret);
1539
1540 static noinline void __unlock_for_delalloc(struct inode *inode,
1541                                            struct page *locked_page,
1542                                            u64 start, u64 end)
1543 {
1544         unsigned long index = start >> PAGE_SHIFT;
1545         unsigned long end_index = end >> PAGE_SHIFT;
1546
1547         ASSERT(locked_page);
1548         if (index == locked_page->index && end_index == index)
1549                 return;
1550
1551         __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1552                                PAGE_UNLOCK, NULL);
1553 }
1554
1555 static noinline int lock_delalloc_pages(struct inode *inode,
1556                                         struct page *locked_page,
1557                                         u64 delalloc_start,
1558                                         u64 delalloc_end)
1559 {
1560         unsigned long index = delalloc_start >> PAGE_SHIFT;
1561         unsigned long index_ret = index;
1562         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1563         int ret;
1564
1565         ASSERT(locked_page);
1566         if (index == locked_page->index && index == end_index)
1567                 return 0;
1568
1569         ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1570                                      end_index, PAGE_LOCK, &index_ret);
1571         if (ret == -EAGAIN)
1572                 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1573                                       (u64)index_ret << PAGE_SHIFT);
1574         return ret;
1575 }
1576
1577 /*
1578  * find a contiguous range of bytes in the file marked as delalloc, not
1579  * more than 'max_bytes'.  start and end are used to return the range,
1580  *
1581  * 1 is returned if we find something, 0 if nothing was in the tree
1582  */
1583 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1584                                     struct extent_io_tree *tree,
1585                                     struct page *locked_page, u64 *start,
1586                                     u64 *end, u64 max_bytes)
1587 {
1588         u64 delalloc_start;
1589         u64 delalloc_end;
1590         u64 found;
1591         struct extent_state *cached_state = NULL;
1592         int ret;
1593         int loops = 0;
1594
1595 again:
1596         /* step one, find a bunch of delalloc bytes starting at start */
1597         delalloc_start = *start;
1598         delalloc_end = 0;
1599         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1600                                     max_bytes, &cached_state);
1601         if (!found || delalloc_end <= *start) {
1602                 *start = delalloc_start;
1603                 *end = delalloc_end;
1604                 free_extent_state(cached_state);
1605                 return 0;
1606         }
1607
1608         /*
1609          * start comes from the offset of locked_page.  We have to lock
1610          * pages in order, so we can't process delalloc bytes before
1611          * locked_page
1612          */
1613         if (delalloc_start < *start)
1614                 delalloc_start = *start;
1615
1616         /*
1617          * make sure to limit the number of pages we try to lock down
1618          */
1619         if (delalloc_end + 1 - delalloc_start > max_bytes)
1620                 delalloc_end = delalloc_start + max_bytes - 1;
1621
1622         /* step two, lock all the pages after the page that has start */
1623         ret = lock_delalloc_pages(inode, locked_page,
1624                                   delalloc_start, delalloc_end);
1625         if (ret == -EAGAIN) {
1626                 /* some of the pages are gone, lets avoid looping by
1627                  * shortening the size of the delalloc range we're searching
1628                  */
1629                 free_extent_state(cached_state);
1630                 cached_state = NULL;
1631                 if (!loops) {
1632                         max_bytes = PAGE_SIZE;
1633                         loops = 1;
1634                         goto again;
1635                 } else {
1636                         found = 0;
1637                         goto out_failed;
1638                 }
1639         }
1640         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1641
1642         /* step three, lock the state bits for the whole range */
1643         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1644
1645         /* then test to make sure it is all still delalloc */
1646         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1647                              EXTENT_DELALLOC, 1, cached_state);
1648         if (!ret) {
1649                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1650                                      &cached_state, GFP_NOFS);
1651                 __unlock_for_delalloc(inode, locked_page,
1652                               delalloc_start, delalloc_end);
1653                 cond_resched();
1654                 goto again;
1655         }
1656         free_extent_state(cached_state);
1657         *start = delalloc_start;
1658         *end = delalloc_end;
1659 out_failed:
1660         return found;
1661 }
1662
1663 static int __process_pages_contig(struct address_space *mapping,
1664                                   struct page *locked_page,
1665                                   pgoff_t start_index, pgoff_t end_index,
1666                                   unsigned long page_ops, pgoff_t *index_ret)
1667 {
1668         unsigned long nr_pages = end_index - start_index + 1;
1669         unsigned long pages_locked = 0;
1670         pgoff_t index = start_index;
1671         struct page *pages[16];
1672         unsigned ret;
1673         int err = 0;
1674         int i;
1675
1676         if (page_ops & PAGE_LOCK) {
1677                 ASSERT(page_ops == PAGE_LOCK);
1678                 ASSERT(index_ret && *index_ret == start_index);
1679         }
1680
1681         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1682                 mapping_set_error(mapping, -EIO);
1683
1684         while (nr_pages > 0) {
1685                 ret = find_get_pages_contig(mapping, index,
1686                                      min_t(unsigned long,
1687                                      nr_pages, ARRAY_SIZE(pages)), pages);
1688                 if (ret == 0) {
1689                         /*
1690                          * Only if we're going to lock these pages,
1691                          * can we find nothing at @index.
1692                          */
1693                         ASSERT(page_ops & PAGE_LOCK);
1694                         err = -EAGAIN;
1695                         goto out;
1696                 }
1697
1698                 for (i = 0; i < ret; i++) {
1699                         if (page_ops & PAGE_SET_PRIVATE2)
1700                                 SetPagePrivate2(pages[i]);
1701
1702                         if (pages[i] == locked_page) {
1703                                 put_page(pages[i]);
1704                                 pages_locked++;
1705                                 continue;
1706                         }
1707                         if (page_ops & PAGE_CLEAR_DIRTY)
1708                                 clear_page_dirty_for_io(pages[i]);
1709                         if (page_ops & PAGE_SET_WRITEBACK)
1710                                 set_page_writeback(pages[i]);
1711                         if (page_ops & PAGE_SET_ERROR)
1712                                 SetPageError(pages[i]);
1713                         if (page_ops & PAGE_END_WRITEBACK)
1714                                 end_page_writeback(pages[i]);
1715                         if (page_ops & PAGE_UNLOCK)
1716                                 unlock_page(pages[i]);
1717                         if (page_ops & PAGE_LOCK) {
1718                                 lock_page(pages[i]);
1719                                 if (!PageDirty(pages[i]) ||
1720                                     pages[i]->mapping != mapping) {
1721                                         unlock_page(pages[i]);
1722                                         put_page(pages[i]);
1723                                         err = -EAGAIN;
1724                                         goto out;
1725                                 }
1726                         }
1727                         put_page(pages[i]);
1728                         pages_locked++;
1729                 }
1730                 nr_pages -= ret;
1731                 index += ret;
1732                 cond_resched();
1733         }
1734 out:
1735         if (err && index_ret)
1736                 *index_ret = start_index + pages_locked - 1;
1737         return err;
1738 }
1739
1740 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1741                                  u64 delalloc_end, struct page *locked_page,
1742                                  unsigned clear_bits,
1743                                  unsigned long page_ops)
1744 {
1745         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1746                          NULL);
1747
1748         __process_pages_contig(inode->i_mapping, locked_page,
1749                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1750                                page_ops, NULL);
1751 }
1752
1753 /*
1754  * count the number of bytes in the tree that have a given bit(s)
1755  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1756  * cached.  The total number found is returned.
1757  */
1758 u64 count_range_bits(struct extent_io_tree *tree,
1759                      u64 *start, u64 search_end, u64 max_bytes,
1760                      unsigned bits, int contig)
1761 {
1762         struct rb_node *node;
1763         struct extent_state *state;
1764         u64 cur_start = *start;
1765         u64 total_bytes = 0;
1766         u64 last = 0;
1767         int found = 0;
1768
1769         if (WARN_ON(search_end <= cur_start))
1770                 return 0;
1771
1772         spin_lock(&tree->lock);
1773         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1774                 total_bytes = tree->dirty_bytes;
1775                 goto out;
1776         }
1777         /*
1778          * this search will find all the extents that end after
1779          * our range starts.
1780          */
1781         node = tree_search(tree, cur_start);
1782         if (!node)
1783                 goto out;
1784
1785         while (1) {
1786                 state = rb_entry(node, struct extent_state, rb_node);
1787                 if (state->start > search_end)
1788                         break;
1789                 if (contig && found && state->start > last + 1)
1790                         break;
1791                 if (state->end >= cur_start && (state->state & bits) == bits) {
1792                         total_bytes += min(search_end, state->end) + 1 -
1793                                        max(cur_start, state->start);
1794                         if (total_bytes >= max_bytes)
1795                                 break;
1796                         if (!found) {
1797                                 *start = max(cur_start, state->start);
1798                                 found = 1;
1799                         }
1800                         last = state->end;
1801                 } else if (contig && found) {
1802                         break;
1803                 }
1804                 node = rb_next(node);
1805                 if (!node)
1806                         break;
1807         }
1808 out:
1809         spin_unlock(&tree->lock);
1810         return total_bytes;
1811 }
1812
1813 /*
1814  * set the private field for a given byte offset in the tree.  If there isn't
1815  * an extent_state there already, this does nothing.
1816  */
1817 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1818                 struct io_failure_record *failrec)
1819 {
1820         struct rb_node *node;
1821         struct extent_state *state;
1822         int ret = 0;
1823
1824         spin_lock(&tree->lock);
1825         /*
1826          * this search will find all the extents that end after
1827          * our range starts.
1828          */
1829         node = tree_search(tree, start);
1830         if (!node) {
1831                 ret = -ENOENT;
1832                 goto out;
1833         }
1834         state = rb_entry(node, struct extent_state, rb_node);
1835         if (state->start != start) {
1836                 ret = -ENOENT;
1837                 goto out;
1838         }
1839         state->failrec = failrec;
1840 out:
1841         spin_unlock(&tree->lock);
1842         return ret;
1843 }
1844
1845 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1846                 struct io_failure_record **failrec)
1847 {
1848         struct rb_node *node;
1849         struct extent_state *state;
1850         int ret = 0;
1851
1852         spin_lock(&tree->lock);
1853         /*
1854          * this search will find all the extents that end after
1855          * our range starts.
1856          */
1857         node = tree_search(tree, start);
1858         if (!node) {
1859                 ret = -ENOENT;
1860                 goto out;
1861         }
1862         state = rb_entry(node, struct extent_state, rb_node);
1863         if (state->start != start) {
1864                 ret = -ENOENT;
1865                 goto out;
1866         }
1867         *failrec = state->failrec;
1868 out:
1869         spin_unlock(&tree->lock);
1870         return ret;
1871 }
1872
1873 /*
1874  * searches a range in the state tree for a given mask.
1875  * If 'filled' == 1, this returns 1 only if every extent in the tree
1876  * has the bits set.  Otherwise, 1 is returned if any bit in the
1877  * range is found set.
1878  */
1879 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1880                    unsigned bits, int filled, struct extent_state *cached)
1881 {
1882         struct extent_state *state = NULL;
1883         struct rb_node *node;
1884         int bitset = 0;
1885
1886         spin_lock(&tree->lock);
1887         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1888             cached->end > start)
1889                 node = &cached->rb_node;
1890         else
1891                 node = tree_search(tree, start);
1892         while (node && start <= end) {
1893                 state = rb_entry(node, struct extent_state, rb_node);
1894
1895                 if (filled && state->start > start) {
1896                         bitset = 0;
1897                         break;
1898                 }
1899
1900                 if (state->start > end)
1901                         break;
1902
1903                 if (state->state & bits) {
1904                         bitset = 1;
1905                         if (!filled)
1906                                 break;
1907                 } else if (filled) {
1908                         bitset = 0;
1909                         break;
1910                 }
1911
1912                 if (state->end == (u64)-1)
1913                         break;
1914
1915                 start = state->end + 1;
1916                 if (start > end)
1917                         break;
1918                 node = rb_next(node);
1919                 if (!node) {
1920                         if (filled)
1921                                 bitset = 0;
1922                         break;
1923                 }
1924         }
1925         spin_unlock(&tree->lock);
1926         return bitset;
1927 }
1928
1929 /*
1930  * helper function to set a given page up to date if all the
1931  * extents in the tree for that page are up to date
1932  */
1933 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1934 {
1935         u64 start = page_offset(page);
1936         u64 end = start + PAGE_SIZE - 1;
1937         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1938                 SetPageUptodate(page);
1939 }
1940
1941 int free_io_failure(struct extent_io_tree *failure_tree,
1942                     struct extent_io_tree *io_tree,
1943                     struct io_failure_record *rec)
1944 {
1945         int ret;
1946         int err = 0;
1947
1948         set_state_failrec(failure_tree, rec->start, NULL);
1949         ret = clear_extent_bits(failure_tree, rec->start,
1950                                 rec->start + rec->len - 1,
1951                                 EXTENT_LOCKED | EXTENT_DIRTY);
1952         if (ret)
1953                 err = ret;
1954
1955         ret = clear_extent_bits(io_tree, rec->start,
1956                                 rec->start + rec->len - 1,
1957                                 EXTENT_DAMAGED);
1958         if (ret && !err)
1959                 err = ret;
1960
1961         kfree(rec);
1962         return err;
1963 }
1964
1965 /*
1966  * this bypasses the standard btrfs submit functions deliberately, as
1967  * the standard behavior is to write all copies in a raid setup. here we only
1968  * want to write the one bad copy. so we do the mapping for ourselves and issue
1969  * submit_bio directly.
1970  * to avoid any synchronization issues, wait for the data after writing, which
1971  * actually prevents the read that triggered the error from finishing.
1972  * currently, there can be no more than two copies of every data bit. thus,
1973  * exactly one rewrite is required.
1974  */
1975 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
1976                       u64 length, u64 logical, struct page *page,
1977                       unsigned int pg_offset, int mirror_num)
1978 {
1979         struct bio *bio;
1980         struct btrfs_device *dev;
1981         u64 map_length = 0;
1982         u64 sector;
1983         struct btrfs_bio *bbio = NULL;
1984         int ret;
1985
1986         ASSERT(!(fs_info->sb->s_flags & SB_RDONLY));
1987         BUG_ON(!mirror_num);
1988
1989         bio = btrfs_io_bio_alloc(1);
1990         bio->bi_iter.bi_size = 0;
1991         map_length = length;
1992
1993         /*
1994          * Avoid races with device replace and make sure our bbio has devices
1995          * associated to its stripes that don't go away while we are doing the
1996          * read repair operation.
1997          */
1998         btrfs_bio_counter_inc_blocked(fs_info);
1999         if (btrfs_is_parity_mirror(fs_info, logical, length)) {
2000                 /*
2001                  * Note that we don't use BTRFS_MAP_WRITE because it's supposed
2002                  * to update all raid stripes, but here we just want to correct
2003                  * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
2004                  * stripe's dev and sector.
2005                  */
2006                 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
2007                                       &map_length, &bbio, 0);
2008                 if (ret) {
2009                         btrfs_bio_counter_dec(fs_info);
2010                         bio_put(bio);
2011                         return -EIO;
2012                 }
2013                 ASSERT(bbio->mirror_num == 1);
2014         } else {
2015                 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2016                                       &map_length, &bbio, mirror_num);
2017                 if (ret) {
2018                         btrfs_bio_counter_dec(fs_info);
2019                         bio_put(bio);
2020                         return -EIO;
2021                 }
2022                 BUG_ON(mirror_num != bbio->mirror_num);
2023         }
2024
2025         sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2026         bio->bi_iter.bi_sector = sector;
2027         dev = bbio->stripes[bbio->mirror_num - 1].dev;
2028         btrfs_put_bbio(bbio);
2029         if (!dev || !dev->bdev || !dev->writeable) {
2030                 btrfs_bio_counter_dec(fs_info);
2031                 bio_put(bio);
2032                 return -EIO;
2033         }
2034         bio_set_dev(bio, dev->bdev);
2035         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2036         bio_add_page(bio, page, length, pg_offset);
2037
2038         if (btrfsic_submit_bio_wait(bio)) {
2039                 /* try to remap that extent elsewhere? */
2040                 btrfs_bio_counter_dec(fs_info);
2041                 bio_put(bio);
2042                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2043                 return -EIO;
2044         }
2045
2046         btrfs_info_rl_in_rcu(fs_info,
2047                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2048                                   ino, start,
2049                                   rcu_str_deref(dev->name), sector);
2050         btrfs_bio_counter_dec(fs_info);
2051         bio_put(bio);
2052         return 0;
2053 }
2054
2055 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2056                          struct extent_buffer *eb, int mirror_num)
2057 {
2058         u64 start = eb->start;
2059         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2060         int ret = 0;
2061
2062         if (sb_rdonly(fs_info->sb))
2063                 return -EROFS;
2064
2065         for (i = 0; i < num_pages; i++) {
2066                 struct page *p = eb->pages[i];
2067
2068                 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
2069                                         start - page_offset(p), mirror_num);
2070                 if (ret)
2071                         break;
2072                 start += PAGE_SIZE;
2073         }
2074
2075         return ret;
2076 }
2077
2078 /*
2079  * each time an IO finishes, we do a fast check in the IO failure tree
2080  * to see if we need to process or clean up an io_failure_record
2081  */
2082 int clean_io_failure(struct btrfs_fs_info *fs_info,
2083                      struct extent_io_tree *failure_tree,
2084                      struct extent_io_tree *io_tree, u64 start,
2085                      struct page *page, u64 ino, unsigned int pg_offset)
2086 {
2087         u64 private;
2088         struct io_failure_record *failrec;
2089         struct extent_state *state;
2090         int num_copies;
2091         int ret;
2092
2093         private = 0;
2094         ret = count_range_bits(failure_tree, &private, (u64)-1, 1,
2095                                EXTENT_DIRTY, 0);
2096         if (!ret)
2097                 return 0;
2098
2099         ret = get_state_failrec(failure_tree, start, &failrec);
2100         if (ret)
2101                 return 0;
2102
2103         BUG_ON(!failrec->this_mirror);
2104
2105         if (failrec->in_validation) {
2106                 /* there was no real error, just free the record */
2107                 btrfs_debug(fs_info,
2108                         "clean_io_failure: freeing dummy error at %llu",
2109                         failrec->start);
2110                 goto out;
2111         }
2112         if (sb_rdonly(fs_info->sb))
2113                 goto out;
2114
2115         spin_lock(&io_tree->lock);
2116         state = find_first_extent_bit_state(io_tree,
2117                                             failrec->start,
2118                                             EXTENT_LOCKED);
2119         spin_unlock(&io_tree->lock);
2120
2121         if (state && state->start <= failrec->start &&
2122             state->end >= failrec->start + failrec->len - 1) {
2123                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2124                                               failrec->len);
2125                 if (num_copies > 1)  {
2126                         repair_io_failure(fs_info, ino, start, failrec->len,
2127                                           failrec->logical, page, pg_offset,
2128                                           failrec->failed_mirror);
2129                 }
2130         }
2131
2132 out:
2133         free_io_failure(failure_tree, io_tree, failrec);
2134
2135         return 0;
2136 }
2137
2138 /*
2139  * Can be called when
2140  * - hold extent lock
2141  * - under ordered extent
2142  * - the inode is freeing
2143  */
2144 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2145 {
2146         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2147         struct io_failure_record *failrec;
2148         struct extent_state *state, *next;
2149
2150         if (RB_EMPTY_ROOT(&failure_tree->state))
2151                 return;
2152
2153         spin_lock(&failure_tree->lock);
2154         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2155         while (state) {
2156                 if (state->start > end)
2157                         break;
2158
2159                 ASSERT(state->end <= end);
2160
2161                 next = next_state(state);
2162
2163                 failrec = state->failrec;
2164                 free_extent_state(state);
2165                 kfree(failrec);
2166
2167                 state = next;
2168         }
2169         spin_unlock(&failure_tree->lock);
2170 }
2171
2172 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2173                 struct io_failure_record **failrec_ret)
2174 {
2175         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2176         struct io_failure_record *failrec;
2177         struct extent_map *em;
2178         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2179         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2180         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2181         int ret;
2182         u64 logical;
2183
2184         ret = get_state_failrec(failure_tree, start, &failrec);
2185         if (ret) {
2186                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2187                 if (!failrec)
2188                         return -ENOMEM;
2189
2190                 failrec->start = start;
2191                 failrec->len = end - start + 1;
2192                 failrec->this_mirror = 0;
2193                 failrec->bio_flags = 0;
2194                 failrec->in_validation = 0;
2195
2196                 read_lock(&em_tree->lock);
2197                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2198                 if (!em) {
2199                         read_unlock(&em_tree->lock);
2200                         kfree(failrec);
2201                         return -EIO;
2202                 }
2203
2204                 if (em->start > start || em->start + em->len <= start) {
2205                         free_extent_map(em);
2206                         em = NULL;
2207                 }
2208                 read_unlock(&em_tree->lock);
2209                 if (!em) {
2210                         kfree(failrec);
2211                         return -EIO;
2212                 }
2213
2214                 logical = start - em->start;
2215                 logical = em->block_start + logical;
2216                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2217                         logical = em->block_start;
2218                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2219                         extent_set_compress_type(&failrec->bio_flags,
2220                                                  em->compress_type);
2221                 }
2222
2223                 btrfs_debug(fs_info,
2224                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2225                         logical, start, failrec->len);
2226
2227                 failrec->logical = logical;
2228                 free_extent_map(em);
2229
2230                 /* set the bits in the private failure tree */
2231                 ret = set_extent_bits(failure_tree, start, end,
2232                                         EXTENT_LOCKED | EXTENT_DIRTY);
2233                 if (ret >= 0)
2234                         ret = set_state_failrec(failure_tree, start, failrec);
2235                 /* set the bits in the inode's tree */
2236                 if (ret >= 0)
2237                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2238                 if (ret < 0) {
2239                         kfree(failrec);
2240                         return ret;
2241                 }
2242         } else {
2243                 btrfs_debug(fs_info,
2244                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2245                         failrec->logical, failrec->start, failrec->len,
2246                         failrec->in_validation);
2247                 /*
2248                  * when data can be on disk more than twice, add to failrec here
2249                  * (e.g. with a list for failed_mirror) to make
2250                  * clean_io_failure() clean all those errors at once.
2251                  */
2252         }
2253
2254         *failrec_ret = failrec;
2255
2256         return 0;
2257 }
2258
2259 bool btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2260                            struct io_failure_record *failrec, int failed_mirror)
2261 {
2262         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2263         int num_copies;
2264
2265         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2266         if (num_copies == 1) {
2267                 /*
2268                  * we only have a single copy of the data, so don't bother with
2269                  * all the retry and error correction code that follows. no
2270                  * matter what the error is, it is very likely to persist.
2271                  */
2272                 btrfs_debug(fs_info,
2273                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2274                         num_copies, failrec->this_mirror, failed_mirror);
2275                 return false;
2276         }
2277
2278         /*
2279          * there are two premises:
2280          *      a) deliver good data to the caller
2281          *      b) correct the bad sectors on disk
2282          */
2283         if (failed_bio->bi_vcnt > 1) {
2284                 /*
2285                  * to fulfill b), we need to know the exact failing sectors, as
2286                  * we don't want to rewrite any more than the failed ones. thus,
2287                  * we need separate read requests for the failed bio
2288                  *
2289                  * if the following BUG_ON triggers, our validation request got
2290                  * merged. we need separate requests for our algorithm to work.
2291                  */
2292                 BUG_ON(failrec->in_validation);
2293                 failrec->in_validation = 1;
2294                 failrec->this_mirror = failed_mirror;
2295         } else {
2296                 /*
2297                  * we're ready to fulfill a) and b) alongside. get a good copy
2298                  * of the failed sector and if we succeed, we have setup
2299                  * everything for repair_io_failure to do the rest for us.
2300                  */
2301                 if (failrec->in_validation) {
2302                         BUG_ON(failrec->this_mirror != failed_mirror);
2303                         failrec->in_validation = 0;
2304                         failrec->this_mirror = 0;
2305                 }
2306                 failrec->failed_mirror = failed_mirror;
2307                 failrec->this_mirror++;
2308                 if (failrec->this_mirror == failed_mirror)
2309                         failrec->this_mirror++;
2310         }
2311
2312         if (failrec->this_mirror > num_copies) {
2313                 btrfs_debug(fs_info,
2314                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2315                         num_copies, failrec->this_mirror, failed_mirror);
2316                 return false;
2317         }
2318
2319         return true;
2320 }
2321
2322
2323 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2324                                     struct io_failure_record *failrec,
2325                                     struct page *page, int pg_offset, int icsum,
2326                                     bio_end_io_t *endio_func, void *data)
2327 {
2328         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2329         struct bio *bio;
2330         struct btrfs_io_bio *btrfs_failed_bio;
2331         struct btrfs_io_bio *btrfs_bio;
2332
2333         bio = btrfs_io_bio_alloc(1);
2334         bio->bi_end_io = endio_func;
2335         bio->bi_iter.bi_sector = failrec->logical >> 9;
2336         bio_set_dev(bio, fs_info->fs_devices->latest_bdev);
2337         bio->bi_iter.bi_size = 0;
2338         bio->bi_private = data;
2339
2340         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2341         if (btrfs_failed_bio->csum) {
2342                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2343
2344                 btrfs_bio = btrfs_io_bio(bio);
2345                 btrfs_bio->csum = btrfs_bio->csum_inline;
2346                 icsum *= csum_size;
2347                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2348                        csum_size);
2349         }
2350
2351         bio_add_page(bio, page, failrec->len, pg_offset);
2352
2353         return bio;
2354 }
2355
2356 /*
2357  * this is a generic handler for readpage errors (default
2358  * readpage_io_failed_hook). if other copies exist, read those and write back
2359  * good data to the failed position. does not investigate in remapping the
2360  * failed extent elsewhere, hoping the device will be smart enough to do this as
2361  * needed
2362  */
2363
2364 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2365                               struct page *page, u64 start, u64 end,
2366                               int failed_mirror)
2367 {
2368         struct io_failure_record *failrec;
2369         struct inode *inode = page->mapping->host;
2370         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2371         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2372         struct bio *bio;
2373         int read_mode = 0;
2374         blk_status_t status;
2375         int ret;
2376
2377         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2378
2379         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2380         if (ret)
2381                 return ret;
2382
2383         if (!btrfs_check_repairable(inode, failed_bio, failrec,
2384                                     failed_mirror)) {
2385                 free_io_failure(failure_tree, tree, failrec);
2386                 return -EIO;
2387         }
2388
2389         if (failed_bio->bi_vcnt > 1)
2390                 read_mode |= REQ_FAILFAST_DEV;
2391
2392         phy_offset >>= inode->i_sb->s_blocksize_bits;
2393         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2394                                       start - page_offset(page),
2395                                       (int)phy_offset, failed_bio->bi_end_io,
2396                                       NULL);
2397         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2398
2399         btrfs_debug(btrfs_sb(inode->i_sb),
2400                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2401                 read_mode, failrec->this_mirror, failrec->in_validation);
2402
2403         status = tree->ops->submit_bio_hook(tree->private_data, bio, failrec->this_mirror,
2404                                          failrec->bio_flags, 0);
2405         if (status) {
2406                 free_io_failure(failure_tree, tree, failrec);
2407                 bio_put(bio);
2408                 ret = blk_status_to_errno(status);
2409         }
2410
2411         return ret;
2412 }
2413
2414 /* lots and lots of room for performance fixes in the end_bio funcs */
2415
2416 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2417 {
2418         int uptodate = (err == 0);
2419         struct extent_io_tree *tree;
2420         int ret = 0;
2421
2422         tree = &BTRFS_I(page->mapping->host)->io_tree;
2423
2424         if (tree->ops && tree->ops->writepage_end_io_hook)
2425                 tree->ops->writepage_end_io_hook(page, start, end, NULL,
2426                                 uptodate);
2427
2428         if (!uptodate) {
2429                 ClearPageUptodate(page);
2430                 SetPageError(page);
2431                 ret = err < 0 ? err : -EIO;
2432                 mapping_set_error(page->mapping, ret);
2433         }
2434 }
2435
2436 /*
2437  * after a writepage IO is done, we need to:
2438  * clear the uptodate bits on error
2439  * clear the writeback bits in the extent tree for this IO
2440  * end_page_writeback if the page has no more pending IO
2441  *
2442  * Scheduling is not allowed, so the extent state tree is expected
2443  * to have one and only one object corresponding to this IO.
2444  */
2445 static void end_bio_extent_writepage(struct bio *bio)
2446 {
2447         int error = blk_status_to_errno(bio->bi_status);
2448         struct bio_vec *bvec;
2449         u64 start;
2450         u64 end;
2451         int i;
2452
2453         ASSERT(!bio_flagged(bio, BIO_CLONED));
2454         bio_for_each_segment_all(bvec, bio, i) {
2455                 struct page *page = bvec->bv_page;
2456                 struct inode *inode = page->mapping->host;
2457                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2458
2459                 /* We always issue full-page reads, but if some block
2460                  * in a page fails to read, blk_update_request() will
2461                  * advance bv_offset and adjust bv_len to compensate.
2462                  * Print a warning for nonzero offsets, and an error
2463                  * if they don't add up to a full page.  */
2464                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2465                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2466                                 btrfs_err(fs_info,
2467                                    "partial page write in btrfs with offset %u and length %u",
2468                                         bvec->bv_offset, bvec->bv_len);
2469                         else
2470                                 btrfs_info(fs_info,
2471                                    "incomplete page write in btrfs with offset %u and length %u",
2472                                         bvec->bv_offset, bvec->bv_len);
2473                 }
2474
2475                 start = page_offset(page);
2476                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2477
2478                 end_extent_writepage(page, error, start, end);
2479                 end_page_writeback(page);
2480         }
2481
2482         bio_put(bio);
2483 }
2484
2485 static void
2486 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2487                               int uptodate)
2488 {
2489         struct extent_state *cached = NULL;
2490         u64 end = start + len - 1;
2491
2492         if (uptodate && tree->track_uptodate)
2493                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2494         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2495 }
2496
2497 /*
2498  * after a readpage IO is done, we need to:
2499  * clear the uptodate bits on error
2500  * set the uptodate bits if things worked
2501  * set the page up to date if all extents in the tree are uptodate
2502  * clear the lock bit in the extent tree
2503  * unlock the page if there are no other extents locked for it
2504  *
2505  * Scheduling is not allowed, so the extent state tree is expected
2506  * to have one and only one object corresponding to this IO.
2507  */
2508 static void end_bio_extent_readpage(struct bio *bio)
2509 {
2510         struct bio_vec *bvec;
2511         int uptodate = !bio->bi_status;
2512         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2513         struct extent_io_tree *tree, *failure_tree;
2514         u64 offset = 0;
2515         u64 start;
2516         u64 end;
2517         u64 len;
2518         u64 extent_start = 0;
2519         u64 extent_len = 0;
2520         int mirror;
2521         int ret;
2522         int i;
2523
2524         ASSERT(!bio_flagged(bio, BIO_CLONED));
2525         bio_for_each_segment_all(bvec, bio, i) {
2526                 struct page *page = bvec->bv_page;
2527                 struct inode *inode = page->mapping->host;
2528                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2529
2530                 btrfs_debug(fs_info,
2531                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2532                         (u64)bio->bi_iter.bi_sector, bio->bi_status,
2533                         io_bio->mirror_num);
2534                 tree = &BTRFS_I(inode)->io_tree;
2535                 failure_tree = &BTRFS_I(inode)->io_failure_tree;
2536
2537                 /* We always issue full-page reads, but if some block
2538                  * in a page fails to read, blk_update_request() will
2539                  * advance bv_offset and adjust bv_len to compensate.
2540                  * Print a warning for nonzero offsets, and an error
2541                  * if they don't add up to a full page.  */
2542                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2543                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2544                                 btrfs_err(fs_info,
2545                                         "partial page read in btrfs with offset %u and length %u",
2546                                         bvec->bv_offset, bvec->bv_len);
2547                         else
2548                                 btrfs_info(fs_info,
2549                                         "incomplete page read in btrfs with offset %u and length %u",
2550                                         bvec->bv_offset, bvec->bv_len);
2551                 }
2552
2553                 start = page_offset(page);
2554                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2555                 len = bvec->bv_len;
2556
2557                 mirror = io_bio->mirror_num;
2558                 if (likely(uptodate && tree->ops)) {
2559                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2560                                                               page, start, end,
2561                                                               mirror);
2562                         if (ret)
2563                                 uptodate = 0;
2564                         else
2565                                 clean_io_failure(BTRFS_I(inode)->root->fs_info,
2566                                                  failure_tree, tree, start,
2567                                                  page,
2568                                                  btrfs_ino(BTRFS_I(inode)), 0);
2569                 }
2570
2571                 if (likely(uptodate))
2572                         goto readpage_ok;
2573
2574                 if (tree->ops) {
2575                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2576                         if (ret == -EAGAIN) {
2577                                 /*
2578                                  * Data inode's readpage_io_failed_hook() always
2579                                  * returns -EAGAIN.
2580                                  *
2581                                  * The generic bio_readpage_error handles errors
2582                                  * the following way: If possible, new read
2583                                  * requests are created and submitted and will
2584                                  * end up in end_bio_extent_readpage as well (if
2585                                  * we're lucky, not in the !uptodate case). In
2586                                  * that case it returns 0 and we just go on with
2587                                  * the next page in our bio. If it can't handle
2588                                  * the error it will return -EIO and we remain
2589                                  * responsible for that page.
2590                                  */
2591                                 ret = bio_readpage_error(bio, offset, page,
2592                                                          start, end, mirror);
2593                                 if (ret == 0) {
2594                                         uptodate = !bio->bi_status;
2595                                         offset += len;
2596                                         continue;
2597                                 }
2598                         }
2599
2600                         /*
2601                          * metadata's readpage_io_failed_hook() always returns
2602                          * -EIO and fixes nothing.  -EIO is also returned if
2603                          * data inode error could not be fixed.
2604                          */
2605                         ASSERT(ret == -EIO);
2606                 }
2607 readpage_ok:
2608                 if (likely(uptodate)) {
2609                         loff_t i_size = i_size_read(inode);
2610                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2611                         unsigned off;
2612
2613                         /* Zero out the end if this page straddles i_size */
2614                         off = i_size & (PAGE_SIZE-1);
2615                         if (page->index == end_index && off)
2616                                 zero_user_segment(page, off, PAGE_SIZE);
2617                         SetPageUptodate(page);
2618                 } else {
2619                         ClearPageUptodate(page);
2620                         SetPageError(page);
2621                 }
2622                 unlock_page(page);
2623                 offset += len;
2624
2625                 if (unlikely(!uptodate)) {
2626                         if (extent_len) {
2627                                 endio_readpage_release_extent(tree,
2628                                                               extent_start,
2629                                                               extent_len, 1);
2630                                 extent_start = 0;
2631                                 extent_len = 0;
2632                         }
2633                         endio_readpage_release_extent(tree, start,
2634                                                       end - start + 1, 0);
2635                 } else if (!extent_len) {
2636                         extent_start = start;
2637                         extent_len = end + 1 - start;
2638                 } else if (extent_start + extent_len == start) {
2639                         extent_len += end + 1 - start;
2640                 } else {
2641                         endio_readpage_release_extent(tree, extent_start,
2642                                                       extent_len, uptodate);
2643                         extent_start = start;
2644                         extent_len = end + 1 - start;
2645                 }
2646         }
2647
2648         if (extent_len)
2649                 endio_readpage_release_extent(tree, extent_start, extent_len,
2650                                               uptodate);
2651         if (io_bio->end_io)
2652                 io_bio->end_io(io_bio, blk_status_to_errno(bio->bi_status));
2653         bio_put(bio);
2654 }
2655
2656 /*
2657  * Initialize the members up to but not including 'bio'. Use after allocating a
2658  * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
2659  * 'bio' because use of __GFP_ZERO is not supported.
2660  */
2661 static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
2662 {
2663         memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
2664 }
2665
2666 /*
2667  * The following helpers allocate a bio. As it's backed by a bioset, it'll
2668  * never fail.  We're returning a bio right now but you can call btrfs_io_bio
2669  * for the appropriate container_of magic
2670  */
2671 struct bio *btrfs_bio_alloc(struct block_device *bdev, u64 first_byte)
2672 {
2673         struct bio *bio;
2674
2675         bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, btrfs_bioset);
2676         bio_set_dev(bio, bdev);
2677         bio->bi_iter.bi_sector = first_byte >> 9;
2678         btrfs_io_bio_init(btrfs_io_bio(bio));
2679         return bio;
2680 }
2681
2682 struct bio *btrfs_bio_clone(struct bio *bio)
2683 {
2684         struct btrfs_io_bio *btrfs_bio;
2685         struct bio *new;
2686
2687         /* Bio allocation backed by a bioset does not fail */
2688         new = bio_clone_fast(bio, GFP_NOFS, btrfs_bioset);
2689         btrfs_bio = btrfs_io_bio(new);
2690         btrfs_io_bio_init(btrfs_bio);
2691         btrfs_bio->iter = bio->bi_iter;
2692         return new;
2693 }
2694
2695 struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
2696 {
2697         struct bio *bio;
2698
2699         /* Bio allocation backed by a bioset does not fail */
2700         bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, btrfs_bioset);
2701         btrfs_io_bio_init(btrfs_io_bio(bio));
2702         return bio;
2703 }
2704
2705 struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
2706 {
2707         struct bio *bio;
2708         struct btrfs_io_bio *btrfs_bio;
2709
2710         /* this will never fail when it's backed by a bioset */
2711         bio = bio_clone_fast(orig, GFP_NOFS, btrfs_bioset);
2712         ASSERT(bio);
2713
2714         btrfs_bio = btrfs_io_bio(bio);
2715         btrfs_io_bio_init(btrfs_bio);
2716
2717         bio_trim(bio, offset >> 9, size >> 9);
2718         btrfs_bio->iter = bio->bi_iter;
2719         return bio;
2720 }
2721
2722 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2723                                        unsigned long bio_flags)
2724 {
2725         blk_status_t ret = 0;
2726         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2727         struct page *page = bvec->bv_page;
2728         struct extent_io_tree *tree = bio->bi_private;
2729         u64 start;
2730
2731         start = page_offset(page) + bvec->bv_offset;
2732
2733         bio->bi_private = NULL;
2734         bio_get(bio);
2735
2736         if (tree->ops)
2737                 ret = tree->ops->submit_bio_hook(tree->private_data, bio,
2738                                            mirror_num, bio_flags, start);
2739         else
2740                 btrfsic_submit_bio(bio);
2741
2742         bio_put(bio);
2743         return blk_status_to_errno(ret);
2744 }
2745
2746 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2747                      unsigned long offset, size_t size, struct bio *bio,
2748                      unsigned long bio_flags)
2749 {
2750         int ret = 0;
2751         if (tree->ops)
2752                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2753                                                 bio_flags);
2754         return ret;
2755
2756 }
2757
2758 /*
2759  * @opf:        bio REQ_OP_* and REQ_* flags as one value
2760  */
2761 static int submit_extent_page(unsigned int opf, struct extent_io_tree *tree,
2762                               struct writeback_control *wbc,
2763                               struct page *page, u64 offset,
2764                               size_t size, unsigned long pg_offset,
2765                               struct block_device *bdev,
2766                               struct bio **bio_ret,
2767                               bio_end_io_t end_io_func,
2768                               int mirror_num,
2769                               unsigned long prev_bio_flags,
2770                               unsigned long bio_flags,
2771                               bool force_bio_submit)
2772 {
2773         int ret = 0;
2774         struct bio *bio;
2775         int contig = 0;
2776         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2777         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2778         sector_t sector = offset >> 9;
2779
2780         if (bio_ret && *bio_ret) {
2781                 bio = *bio_ret;
2782                 if (old_compressed)
2783                         contig = bio->bi_iter.bi_sector == sector;
2784                 else
2785                         contig = bio_end_sector(bio) == sector;
2786
2787                 if (prev_bio_flags != bio_flags || !contig ||
2788                     force_bio_submit ||
2789                     merge_bio(tree, page, pg_offset, page_size, bio, bio_flags) ||
2790                     bio_add_page(bio, page, page_size, pg_offset) < page_size) {
2791                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2792                         if (ret < 0) {
2793                                 *bio_ret = NULL;
2794                                 return ret;
2795                         }
2796                         bio = NULL;
2797                 } else {
2798                         if (wbc)
2799                                 wbc_account_io(wbc, page, page_size);
2800                         return 0;
2801                 }
2802         }
2803
2804         bio = btrfs_bio_alloc(bdev, offset);
2805         bio_add_page(bio, page, page_size, pg_offset);
2806         bio->bi_end_io = end_io_func;
2807         bio->bi_private = tree;
2808         bio->bi_write_hint = page->mapping->host->i_write_hint;
2809         bio->bi_opf = opf;
2810         if (wbc) {
2811                 wbc_init_bio(wbc, bio);
2812                 wbc_account_io(wbc, page, page_size);
2813         }
2814
2815         if (bio_ret)
2816                 *bio_ret = bio;
2817         else
2818                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2819
2820         return ret;
2821 }
2822
2823 static void attach_extent_buffer_page(struct extent_buffer *eb,
2824                                       struct page *page)
2825 {
2826         if (!PagePrivate(page)) {
2827                 SetPagePrivate(page);
2828                 get_page(page);
2829                 set_page_private(page, (unsigned long)eb);
2830         } else {
2831                 WARN_ON(page->private != (unsigned long)eb);
2832         }
2833 }
2834
2835 void set_page_extent_mapped(struct page *page)
2836 {
2837         if (!PagePrivate(page)) {
2838                 SetPagePrivate(page);
2839                 get_page(page);
2840                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2841         }
2842 }
2843
2844 static struct extent_map *
2845 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2846                  u64 start, u64 len, get_extent_t *get_extent,
2847                  struct extent_map **em_cached)
2848 {
2849         struct extent_map *em;
2850
2851         if (em_cached && *em_cached) {
2852                 em = *em_cached;
2853                 if (extent_map_in_tree(em) && start >= em->start &&
2854                     start < extent_map_end(em)) {
2855                         refcount_inc(&em->refs);
2856                         return em;
2857                 }
2858
2859                 free_extent_map(em);
2860                 *em_cached = NULL;
2861         }
2862
2863         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2864         if (em_cached && !IS_ERR_OR_NULL(em)) {
2865                 BUG_ON(*em_cached);
2866                 refcount_inc(&em->refs);
2867                 *em_cached = em;
2868         }
2869         return em;
2870 }
2871 /*
2872  * basic readpage implementation.  Locked extent state structs are inserted
2873  * into the tree that are removed when the IO is done (by the end_io
2874  * handlers)
2875  * XXX JDM: This needs looking at to ensure proper page locking
2876  * return 0 on success, otherwise return error
2877  */
2878 static int __do_readpage(struct extent_io_tree *tree,
2879                          struct page *page,
2880                          get_extent_t *get_extent,
2881                          struct extent_map **em_cached,
2882                          struct bio **bio, int mirror_num,
2883                          unsigned long *bio_flags, unsigned int read_flags,
2884                          u64 *prev_em_start)
2885 {
2886         struct inode *inode = page->mapping->host;
2887         u64 start = page_offset(page);
2888         u64 page_end = start + PAGE_SIZE - 1;
2889         u64 end;
2890         u64 cur = start;
2891         u64 extent_offset;
2892         u64 last_byte = i_size_read(inode);
2893         u64 block_start;
2894         u64 cur_end;
2895         struct extent_map *em;
2896         struct block_device *bdev;
2897         int ret = 0;
2898         int nr = 0;
2899         size_t pg_offset = 0;
2900         size_t iosize;
2901         size_t disk_io_size;
2902         size_t blocksize = inode->i_sb->s_blocksize;
2903         unsigned long this_bio_flag = 0;
2904
2905         set_page_extent_mapped(page);
2906
2907         end = page_end;
2908         if (!PageUptodate(page)) {
2909                 if (cleancache_get_page(page) == 0) {
2910                         BUG_ON(blocksize != PAGE_SIZE);
2911                         unlock_extent(tree, start, end);
2912                         goto out;
2913                 }
2914         }
2915
2916         if (page->index == last_byte >> PAGE_SHIFT) {
2917                 char *userpage;
2918                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2919
2920                 if (zero_offset) {
2921                         iosize = PAGE_SIZE - zero_offset;
2922                         userpage = kmap_atomic(page);
2923                         memset(userpage + zero_offset, 0, iosize);
2924                         flush_dcache_page(page);
2925                         kunmap_atomic(userpage);
2926                 }
2927         }
2928         while (cur <= end) {
2929                 bool force_bio_submit = false;
2930                 u64 offset;
2931
2932                 if (cur >= last_byte) {
2933                         char *userpage;
2934                         struct extent_state *cached = NULL;
2935
2936                         iosize = PAGE_SIZE - pg_offset;
2937                         userpage = kmap_atomic(page);
2938                         memset(userpage + pg_offset, 0, iosize);
2939                         flush_dcache_page(page);
2940                         kunmap_atomic(userpage);
2941                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2942                                             &cached, GFP_NOFS);
2943                         unlock_extent_cached(tree, cur,
2944                                              cur + iosize - 1,
2945                                              &cached, GFP_NOFS);
2946                         break;
2947                 }
2948                 em = __get_extent_map(inode, page, pg_offset, cur,
2949                                       end - cur + 1, get_extent, em_cached);
2950                 if (IS_ERR_OR_NULL(em)) {
2951                         SetPageError(page);
2952                         unlock_extent(tree, cur, end);
2953                         break;
2954                 }
2955                 extent_offset = cur - em->start;
2956                 BUG_ON(extent_map_end(em) <= cur);
2957                 BUG_ON(end < cur);
2958
2959                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2960                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2961                         extent_set_compress_type(&this_bio_flag,
2962                                                  em->compress_type);
2963                 }
2964
2965                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2966                 cur_end = min(extent_map_end(em) - 1, end);
2967                 iosize = ALIGN(iosize, blocksize);
2968                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2969                         disk_io_size = em->block_len;
2970                         offset = em->block_start;
2971                 } else {
2972                         offset = em->block_start + extent_offset;
2973                         disk_io_size = iosize;
2974                 }
2975                 bdev = em->bdev;
2976                 block_start = em->block_start;
2977                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2978                         block_start = EXTENT_MAP_HOLE;
2979
2980                 /*
2981                  * If we have a file range that points to a compressed extent
2982                  * and it's followed by a consecutive file range that points to
2983                  * to the same compressed extent (possibly with a different
2984                  * offset and/or length, so it either points to the whole extent
2985                  * or only part of it), we must make sure we do not submit a
2986                  * single bio to populate the pages for the 2 ranges because
2987                  * this makes the compressed extent read zero out the pages
2988                  * belonging to the 2nd range. Imagine the following scenario:
2989                  *
2990                  *  File layout
2991                  *  [0 - 8K]                     [8K - 24K]
2992                  *    |                               |
2993                  *    |                               |
2994                  * points to extent X,         points to extent X,
2995                  * offset 4K, length of 8K     offset 0, length 16K
2996                  *
2997                  * [extent X, compressed length = 4K uncompressed length = 16K]
2998                  *
2999                  * If the bio to read the compressed extent covers both ranges,
3000                  * it will decompress extent X into the pages belonging to the
3001                  * first range and then it will stop, zeroing out the remaining
3002                  * pages that belong to the other range that points to extent X.
3003                  * So here we make sure we submit 2 bios, one for the first
3004                  * range and another one for the third range. Both will target
3005                  * the same physical extent from disk, but we can't currently
3006                  * make the compressed bio endio callback populate the pages
3007                  * for both ranges because each compressed bio is tightly
3008                  * coupled with a single extent map, and each range can have
3009                  * an extent map with a different offset value relative to the
3010                  * uncompressed data of our extent and different lengths. This
3011                  * is a corner case so we prioritize correctness over
3012                  * non-optimal behavior (submitting 2 bios for the same extent).
3013                  */
3014                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3015                     prev_em_start && *prev_em_start != (u64)-1 &&
3016                     *prev_em_start != em->orig_start)
3017                         force_bio_submit = true;
3018
3019                 if (prev_em_start)
3020                         *prev_em_start = em->orig_start;
3021
3022                 free_extent_map(em);
3023                 em = NULL;
3024
3025                 /* we've found a hole, just zero and go on */
3026                 if (block_start == EXTENT_MAP_HOLE) {
3027                         char *userpage;
3028                         struct extent_state *cached = NULL;
3029
3030                         userpage = kmap_atomic(page);
3031                         memset(userpage + pg_offset, 0, iosize);
3032                         flush_dcache_page(page);
3033                         kunmap_atomic(userpage);
3034
3035                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3036                                             &cached, GFP_NOFS);
3037                         unlock_extent_cached(tree, cur,
3038                                              cur + iosize - 1,
3039                                              &cached, GFP_NOFS);
3040                         cur = cur + iosize;
3041                         pg_offset += iosize;
3042                         continue;
3043                 }
3044                 /* the get_extent function already copied into the page */
3045                 if (test_range_bit(tree, cur, cur_end,
3046                                    EXTENT_UPTODATE, 1, NULL)) {
3047                         check_page_uptodate(tree, page);
3048                         unlock_extent(tree, cur, cur + iosize - 1);
3049                         cur = cur + iosize;
3050                         pg_offset += iosize;
3051                         continue;
3052                 }
3053                 /* we have an inline extent but it didn't get marked up
3054                  * to date.  Error out
3055                  */
3056                 if (block_start == EXTENT_MAP_INLINE) {
3057                         SetPageError(page);
3058                         unlock_extent(tree, cur, cur + iosize - 1);
3059                         cur = cur + iosize;
3060                         pg_offset += iosize;
3061                         continue;
3062                 }
3063
3064                 ret = submit_extent_page(REQ_OP_READ | read_flags, tree, NULL,
3065                                          page, offset, disk_io_size,
3066                                          pg_offset, bdev, bio,
3067                                          end_bio_extent_readpage, mirror_num,
3068                                          *bio_flags,
3069                                          this_bio_flag,
3070                                          force_bio_submit);
3071                 if (!ret) {
3072                         nr++;
3073                         *bio_flags = this_bio_flag;
3074                 } else {
3075                         SetPageError(page);
3076                         unlock_extent(tree, cur, cur + iosize - 1);
3077                         goto out;
3078                 }
3079                 cur = cur + iosize;
3080                 pg_offset += iosize;
3081         }
3082 out:
3083         if (!nr) {
3084                 if (!PageError(page))
3085                         SetPageUptodate(page);
3086                 unlock_page(page);
3087         }
3088         return ret;
3089 }
3090
3091 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3092                                              struct page *pages[], int nr_pages,
3093                                              u64 start, u64 end,
3094                                              struct extent_map **em_cached,
3095                                              struct bio **bio,
3096                                              unsigned long *bio_flags,
3097                                              u64 *prev_em_start)
3098 {
3099         struct inode *inode;
3100         struct btrfs_ordered_extent *ordered;
3101         int index;
3102
3103         inode = pages[0]->mapping->host;
3104         while (1) {
3105                 lock_extent(tree, start, end);
3106                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3107                                                      end - start + 1);
3108                 if (!ordered)
3109                         break;
3110                 unlock_extent(tree, start, end);
3111                 btrfs_start_ordered_extent(inode, ordered, 1);
3112                 btrfs_put_ordered_extent(ordered);
3113         }
3114
3115         for (index = 0; index < nr_pages; index++) {
3116                 __do_readpage(tree, pages[index], btrfs_get_extent, em_cached,
3117                                 bio, 0, bio_flags, 0, prev_em_start);
3118                 put_page(pages[index]);
3119         }
3120 }
3121
3122 static void __extent_readpages(struct extent_io_tree *tree,
3123                                struct page *pages[],
3124                                int nr_pages,
3125                                struct extent_map **em_cached,
3126                                struct bio **bio, 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, em_cached,
3147                                                   bio, 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, em_cached, bio,
3159                                           bio_flags, prev_em_start);
3160 }
3161
3162 static int __extent_read_full_page(struct extent_io_tree *tree,
3163                                    struct page *page,
3164                                    get_extent_t *get_extent,
3165                                    struct bio **bio, int mirror_num,
3166                                    unsigned long *bio_flags,
3167                                    unsigned 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, wbc);
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                                  unsigned 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         struct extent_map *em;
3325         struct block_device *bdev;
3326         size_t pg_offset = 0;
3327         size_t blocksize;
3328         int ret = 0;
3329         int nr = 0;
3330         bool compressed;
3331
3332         if (tree->ops && tree->ops->writepage_start_hook) {
3333                 ret = tree->ops->writepage_start_hook(page, start,
3334                                                       page_end);
3335                 if (ret) {
3336                         /* Fixup worker will requeue */
3337                         if (ret == -EBUSY)
3338                                 wbc->pages_skipped++;
3339                         else
3340                                 redirty_page_for_writepage(wbc, page);
3341
3342                         update_nr_written(wbc, nr_written);
3343                         unlock_page(page);
3344                         return 1;
3345                 }
3346         }
3347
3348         /*
3349          * we don't want to touch the inode after unlocking the page,
3350          * so we update the mapping writeback index now
3351          */
3352         update_nr_written(wbc, nr_written + 1);
3353
3354         end = page_end;
3355         if (i_size <= start) {
3356                 if (tree->ops && tree->ops->writepage_end_io_hook)
3357                         tree->ops->writepage_end_io_hook(page, start,
3358                                                          page_end, NULL, 1);
3359                 goto done;
3360         }
3361
3362         blocksize = inode->i_sb->s_blocksize;
3363
3364         while (cur <= end) {
3365                 u64 em_end;
3366                 u64 offset;
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 = btrfs_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                 offset = em->block_start + extent_offset;
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, offset, 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         unsigned int write_flags = 0;
3469         unsigned long nr_written = 0;
3470
3471         write_flags = wbc_to_write_flags(wbc);
3472
3473         trace___extent_writepage(page, inode, wbc);
3474
3475         WARN_ON(!PageLocked(page));
3476
3477         ClearPageError(page);
3478
3479         pg_offset = i_size & (PAGE_SIZE - 1);
3480         if (page->index > end_index ||
3481            (page->index == end_index && !pg_offset)) {
3482                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3483                 unlock_page(page);
3484                 return 0;
3485         }
3486
3487         if (page->index == end_index) {
3488                 char *userpage;
3489
3490                 userpage = kmap_atomic(page);
3491                 memset(userpage + pg_offset, 0,
3492                        PAGE_SIZE - pg_offset);
3493                 kunmap_atomic(userpage);
3494                 flush_dcache_page(page);
3495         }
3496
3497         pg_offset = 0;
3498
3499         set_page_extent_mapped(page);
3500
3501         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3502         if (ret == 1)
3503                 goto done_unlocked;
3504         if (ret)
3505                 goto done;
3506
3507         ret = __extent_writepage_io(inode, page, wbc, epd,
3508                                     i_size, nr_written, write_flags, &nr);
3509         if (ret == 1)
3510                 goto done_unlocked;
3511
3512 done:
3513         if (nr == 0) {
3514                 /* make sure the mapping tag for page dirty gets cleared */
3515                 set_page_writeback(page);
3516                 end_page_writeback(page);
3517         }
3518         if (PageError(page)) {
3519                 ret = ret < 0 ? ret : -EIO;
3520                 end_extent_writepage(page, ret, start, page_end);
3521         }
3522         unlock_page(page);
3523         return ret;
3524
3525 done_unlocked:
3526         return 0;
3527 }
3528
3529 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3530 {
3531         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3532                        TASK_UNINTERRUPTIBLE);
3533 }
3534
3535 static noinline_for_stack int
3536 lock_extent_buffer_for_io(struct extent_buffer *eb,
3537                           struct btrfs_fs_info *fs_info,
3538                           struct extent_page_data *epd)
3539 {
3540         unsigned long i, num_pages;
3541         int flush = 0;
3542         int ret = 0;
3543
3544         if (!btrfs_try_tree_write_lock(eb)) {
3545                 flush = 1;
3546                 flush_write_bio(epd);
3547                 btrfs_tree_lock(eb);
3548         }
3549
3550         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3551                 btrfs_tree_unlock(eb);
3552                 if (!epd->sync_io)
3553                         return 0;
3554                 if (!flush) {
3555                         flush_write_bio(epd);
3556                         flush = 1;
3557                 }
3558                 while (1) {
3559                         wait_on_extent_buffer_writeback(eb);
3560                         btrfs_tree_lock(eb);
3561                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3562                                 break;
3563                         btrfs_tree_unlock(eb);
3564                 }
3565         }
3566
3567         /*
3568          * We need to do this to prevent races in people who check if the eb is
3569          * under IO since we can end up having no IO bits set for a short period
3570          * of time.
3571          */
3572         spin_lock(&eb->refs_lock);
3573         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3574                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3575                 spin_unlock(&eb->refs_lock);
3576                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3577                 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
3578                                          -eb->len,
3579                                          fs_info->dirty_metadata_batch);
3580                 ret = 1;
3581         } else {
3582                 spin_unlock(&eb->refs_lock);
3583         }
3584
3585         btrfs_tree_unlock(eb);
3586
3587         if (!ret)
3588                 return ret;
3589
3590         num_pages = num_extent_pages(eb->start, eb->len);
3591         for (i = 0; i < num_pages; i++) {
3592                 struct page *p = eb->pages[i];
3593
3594                 if (!trylock_page(p)) {
3595                         if (!flush) {
3596                                 flush_write_bio(epd);
3597                                 flush = 1;
3598                         }
3599                         lock_page(p);
3600                 }
3601         }
3602
3603         return ret;
3604 }
3605
3606 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3607 {
3608         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3609         smp_mb__after_atomic();
3610         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3611 }
3612
3613 static void set_btree_ioerr(struct page *page)
3614 {
3615         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3616
3617         SetPageError(page);
3618         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3619                 return;
3620
3621         /*
3622          * If writeback for a btree extent that doesn't belong to a log tree
3623          * failed, increment the counter transaction->eb_write_errors.
3624          * We do this because while the transaction is running and before it's
3625          * committing (when we call filemap_fdata[write|wait]_range against
3626          * the btree inode), we might have
3627          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3628          * returns an error or an error happens during writeback, when we're
3629          * committing the transaction we wouldn't know about it, since the pages
3630          * can be no longer dirty nor marked anymore for writeback (if a
3631          * subsequent modification to the extent buffer didn't happen before the
3632          * transaction commit), which makes filemap_fdata[write|wait]_range not
3633          * able to find the pages tagged with SetPageError at transaction
3634          * commit time. So if this happens we must abort the transaction,
3635          * otherwise we commit a super block with btree roots that point to
3636          * btree nodes/leafs whose content on disk is invalid - either garbage
3637          * or the content of some node/leaf from a past generation that got
3638          * cowed or deleted and is no longer valid.
3639          *
3640          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3641          * not be enough - we need to distinguish between log tree extents vs
3642          * non-log tree extents, and the next filemap_fdatawait_range() call
3643          * will catch and clear such errors in the mapping - and that call might
3644          * be from a log sync and not from a transaction commit. Also, checking
3645          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3646          * not done and would not be reliable - the eb might have been released
3647          * from memory and reading it back again means that flag would not be
3648          * set (since it's a runtime flag, not persisted on disk).
3649          *
3650          * Using the flags below in the btree inode also makes us achieve the
3651          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3652          * writeback for all dirty pages and before filemap_fdatawait_range()
3653          * is called, the writeback for all dirty pages had already finished
3654          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3655          * filemap_fdatawait_range() would return success, as it could not know
3656          * that writeback errors happened (the pages were no longer tagged for
3657          * writeback).
3658          */
3659         switch (eb->log_index) {
3660         case -1:
3661                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3662                 break;
3663         case 0:
3664                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3665                 break;
3666         case 1:
3667                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3668                 break;
3669         default:
3670                 BUG(); /* unexpected, logic error */
3671         }
3672 }
3673
3674 static void end_bio_extent_buffer_writepage(struct bio *bio)
3675 {
3676         struct bio_vec *bvec;
3677         struct extent_buffer *eb;
3678         int i, done;
3679
3680         ASSERT(!bio_flagged(bio, BIO_CLONED));
3681         bio_for_each_segment_all(bvec, bio, i) {
3682                 struct page *page = bvec->bv_page;
3683
3684                 eb = (struct extent_buffer *)page->private;
3685                 BUG_ON(!eb);
3686                 done = atomic_dec_and_test(&eb->io_pages);
3687
3688                 if (bio->bi_status ||
3689                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3690                         ClearPageUptodate(page);
3691                         set_btree_ioerr(page);
3692                 }
3693
3694                 end_page_writeback(page);
3695
3696                 if (!done)
3697                         continue;
3698
3699                 end_extent_buffer_writeback(eb);
3700         }
3701
3702         bio_put(bio);
3703 }
3704
3705 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3706                         struct btrfs_fs_info *fs_info,
3707                         struct writeback_control *wbc,
3708                         struct extent_page_data *epd)
3709 {
3710         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3711         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3712         u64 offset = eb->start;
3713         u32 nritems;
3714         unsigned long i, num_pages;
3715         unsigned long start, end;
3716         unsigned int write_flags = wbc_to_write_flags(wbc) | REQ_META;
3717         int ret = 0;
3718
3719         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3720         num_pages = num_extent_pages(eb->start, eb->len);
3721         atomic_set(&eb->io_pages, num_pages);
3722
3723         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3724         nritems = btrfs_header_nritems(eb);
3725         if (btrfs_header_level(eb) > 0) {
3726                 end = btrfs_node_key_ptr_offset(nritems);
3727
3728                 memzero_extent_buffer(eb, end, eb->len - end);
3729         } else {
3730                 /*
3731                  * leaf:
3732                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3733                  */
3734                 start = btrfs_item_nr_offset(nritems);
3735                 end = BTRFS_LEAF_DATA_OFFSET + leaf_data_end(fs_info, eb);
3736                 memzero_extent_buffer(eb, start, end - start);
3737         }
3738
3739         for (i = 0; i < num_pages; i++) {
3740                 struct page *p = eb->pages[i];
3741
3742                 clear_page_dirty_for_io(p);
3743                 set_page_writeback(p);
3744                 ret = submit_extent_page(REQ_OP_WRITE | write_flags, tree, wbc,
3745                                          p, offset, PAGE_SIZE, 0, bdev,
3746                                          &epd->bio,
3747                                          end_bio_extent_buffer_writepage,
3748                                          0, 0, 0, false);
3749                 if (ret) {
3750                         set_btree_ioerr(p);
3751                         if (PageWriteback(p))
3752                                 end_page_writeback(p);
3753                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3754                                 end_extent_buffer_writeback(eb);
3755                         ret = -EIO;
3756                         break;
3757                 }
3758                 offset += PAGE_SIZE;
3759                 update_nr_written(wbc, 1);
3760                 unlock_page(p);
3761         }
3762
3763         if (unlikely(ret)) {
3764                 for (; i < num_pages; i++) {
3765                         struct page *p = eb->pages[i];
3766                         clear_page_dirty_for_io(p);
3767                         unlock_page(p);
3768                 }
3769         }
3770
3771         return ret;
3772 }
3773
3774 int btree_write_cache_pages(struct address_space *mapping,
3775                                    struct writeback_control *wbc)
3776 {
3777         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3778         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3779         struct extent_buffer *eb, *prev_eb = NULL;
3780         struct extent_page_data epd = {
3781                 .bio = NULL,
3782                 .tree = tree,
3783                 .extent_locked = 0,
3784                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3785         };
3786         int ret = 0;
3787         int done = 0;
3788         int nr_to_write_done = 0;
3789         struct pagevec pvec;
3790         int nr_pages;
3791         pgoff_t index;
3792         pgoff_t end;            /* Inclusive */
3793         int scanned = 0;
3794         int tag;
3795
3796         pagevec_init(&pvec);
3797         if (wbc->range_cyclic) {
3798                 index = mapping->writeback_index; /* Start from prev offset */
3799                 end = -1;
3800         } else {
3801                 index = wbc->range_start >> PAGE_SHIFT;
3802                 end = wbc->range_end >> PAGE_SHIFT;
3803                 scanned = 1;
3804         }
3805         if (wbc->sync_mode == WB_SYNC_ALL)
3806                 tag = PAGECACHE_TAG_TOWRITE;
3807         else
3808                 tag = PAGECACHE_TAG_DIRTY;
3809 retry:
3810         if (wbc->sync_mode == WB_SYNC_ALL)
3811                 tag_pages_for_writeback(mapping, index, end);
3812         while (!done && !nr_to_write_done && (index <= end) &&
3813                (nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
3814                         tag))) {
3815                 unsigned i;
3816
3817                 scanned = 1;
3818                 for (i = 0; i < nr_pages; i++) {
3819                         struct page *page = pvec.pages[i];
3820
3821                         if (!PagePrivate(page))
3822                                 continue;
3823
3824                         spin_lock(&mapping->private_lock);
3825                         if (!PagePrivate(page)) {
3826                                 spin_unlock(&mapping->private_lock);
3827                                 continue;
3828                         }
3829
3830                         eb = (struct extent_buffer *)page->private;
3831
3832                         /*
3833                          * Shouldn't happen and normally this would be a BUG_ON
3834                          * but no sense in crashing the users box for something
3835                          * we can survive anyway.
3836                          */
3837                         if (WARN_ON(!eb)) {
3838                                 spin_unlock(&mapping->private_lock);
3839                                 continue;
3840                         }
3841
3842                         if (eb == prev_eb) {
3843                                 spin_unlock(&mapping->private_lock);
3844                                 continue;
3845                         }
3846
3847                         ret = atomic_inc_not_zero(&eb->refs);
3848                         spin_unlock(&mapping->private_lock);
3849                         if (!ret)
3850                                 continue;
3851
3852                         prev_eb = eb;
3853                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3854                         if (!ret) {
3855                                 free_extent_buffer(eb);
3856                                 continue;
3857                         }
3858
3859                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3860                         if (ret) {
3861                                 done = 1;
3862                                 free_extent_buffer(eb);
3863                                 break;
3864                         }
3865                         free_extent_buffer(eb);
3866
3867                         /*
3868                          * the filesystem may choose to bump up nr_to_write.
3869                          * We have to make sure to honor the new nr_to_write
3870                          * at any time
3871                          */
3872                         nr_to_write_done = wbc->nr_to_write <= 0;
3873                 }
3874                 pagevec_release(&pvec);
3875                 cond_resched();
3876         }
3877         if (!scanned && !done) {
3878                 /*
3879                  * We hit the last page and there is more work to be done: wrap
3880                  * back to the start of the file
3881                  */
3882                 scanned = 1;
3883                 index = 0;
3884                 goto retry;
3885         }
3886         flush_write_bio(&epd);
3887         return ret;
3888 }
3889
3890 /**
3891  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3892  * @mapping: address space structure to write
3893  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3894  * @writepage: function called for each page
3895  * @data: data passed to writepage function
3896  *
3897  * If a page is already under I/O, write_cache_pages() skips it, even
3898  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3899  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3900  * and msync() need to guarantee that all the data which was dirty at the time
3901  * the call was made get new I/O started against them.  If wbc->sync_mode is
3902  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3903  * existing IO to complete.
3904  */
3905 static int extent_write_cache_pages(struct address_space *mapping,
3906                              struct writeback_control *wbc,
3907                              writepage_t writepage, void *data,
3908                              void (*flush_fn)(void *))
3909 {
3910         struct inode *inode = mapping->host;
3911         int ret = 0;
3912         int done = 0;
3913         int nr_to_write_done = 0;
3914         struct pagevec pvec;
3915         int nr_pages;
3916         pgoff_t index;
3917         pgoff_t end;            /* Inclusive */
3918         pgoff_t done_index;
3919         int range_whole = 0;
3920         int scanned = 0;
3921         int tag;
3922
3923         /*
3924          * We have to hold onto the inode so that ordered extents can do their
3925          * work when the IO finishes.  The alternative to this is failing to add
3926          * an ordered extent if the igrab() fails there and that is a huge pain
3927          * to deal with, so instead just hold onto the inode throughout the
3928          * writepages operation.  If it fails here we are freeing up the inode
3929          * anyway and we'd rather not waste our time writing out stuff that is
3930          * going to be truncated anyway.
3931          */
3932         if (!igrab(inode))
3933                 return 0;
3934
3935         pagevec_init(&pvec);
3936         if (wbc->range_cyclic) {
3937                 index = mapping->writeback_index; /* Start from prev offset */
3938                 end = -1;
3939         } else {
3940                 index = wbc->range_start >> PAGE_SHIFT;
3941                 end = wbc->range_end >> PAGE_SHIFT;
3942                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3943                         range_whole = 1;
3944                 scanned = 1;
3945         }
3946         if (wbc->sync_mode == WB_SYNC_ALL)
3947                 tag = PAGECACHE_TAG_TOWRITE;
3948         else
3949                 tag = PAGECACHE_TAG_DIRTY;
3950 retry:
3951         if (wbc->sync_mode == WB_SYNC_ALL)
3952                 tag_pages_for_writeback(mapping, index, end);
3953         done_index = index;
3954         while (!done && !nr_to_write_done && (index <= end) &&
3955                         (nr_pages = pagevec_lookup_range_tag(&pvec, mapping,
3956                                                 &index, end, tag))) {
3957                 unsigned i;
3958
3959                 scanned = 1;
3960                 for (i = 0; i < nr_pages; i++) {
3961                         struct page *page = pvec.pages[i];
3962
3963                         done_index = page->index;
3964                         /*
3965                          * At this point we hold neither mapping->tree_lock nor
3966                          * lock on the page itself: the page may be truncated or
3967                          * invalidated (changing page->mapping to NULL), or even
3968                          * swizzled back from swapper_space to tmpfs file
3969                          * mapping
3970                          */
3971                         if (!trylock_page(page)) {
3972                                 flush_fn(data);
3973                                 lock_page(page);
3974                         }
3975
3976                         if (unlikely(page->mapping != mapping)) {
3977                                 unlock_page(page);
3978                                 continue;
3979                         }
3980
3981                         if (wbc->sync_mode != WB_SYNC_NONE) {
3982                                 if (PageWriteback(page))
3983                                         flush_fn(data);
3984                                 wait_on_page_writeback(page);
3985                         }
3986
3987                         if (PageWriteback(page) ||
3988                             !clear_page_dirty_for_io(page)) {
3989                                 unlock_page(page);
3990                                 continue;
3991                         }
3992
3993                         ret = (*writepage)(page, wbc, data);
3994
3995                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3996                                 unlock_page(page);
3997                                 ret = 0;
3998                         }
3999                         if (ret < 0) {
4000                                 /*
4001                                  * done_index is set past this page,
4002                                  * so media errors will not choke
4003                                  * background writeout for the entire
4004                                  * file. This has consequences for
4005                                  * range_cyclic semantics (ie. it may
4006                                  * not be suitable for data integrity
4007                                  * writeout).
4008                                  */
4009                                 done_index = page->index + 1;
4010                                 done = 1;
4011                                 break;
4012                         }
4013
4014                         /*
4015                          * the filesystem may choose to bump up nr_to_write.
4016                          * We have to make sure to honor the new nr_to_write
4017                          * at any time
4018                          */
4019                         nr_to_write_done = wbc->nr_to_write <= 0;
4020                 }
4021                 pagevec_release(&pvec);
4022                 cond_resched();
4023         }
4024         if (!scanned && !done) {
4025                 /*
4026                  * We hit the last page and there is more work to be done: wrap
4027                  * back to the start of the file
4028                  */
4029                 scanned = 1;
4030                 index = 0;
4031                 goto retry;
4032         }
4033
4034         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4035                 mapping->writeback_index = done_index;
4036
4037         btrfs_add_delayed_iput(inode);
4038         return ret;
4039 }
4040
4041 static void flush_epd_write_bio(struct extent_page_data *epd)
4042 {
4043         if (epd->bio) {
4044                 int ret;
4045
4046                 ret = submit_one_bio(epd->bio, 0, 0);
4047                 BUG_ON(ret < 0); /* -ENOMEM */
4048                 epd->bio = NULL;
4049         }
4050 }
4051
4052 static noinline void flush_write_bio(void *data)
4053 {
4054         struct extent_page_data *epd = data;
4055         flush_epd_write_bio(epd);
4056 }
4057
4058 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4059                           struct writeback_control *wbc)
4060 {
4061         int ret;
4062         struct extent_page_data epd = {
4063                 .bio = NULL,
4064                 .tree = tree,
4065                 .extent_locked = 0,
4066                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4067         };
4068
4069         ret = __extent_writepage(page, wbc, &epd);
4070
4071         flush_epd_write_bio(&epd);
4072         return ret;
4073 }
4074
4075 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4076                               u64 start, u64 end, int mode)
4077 {
4078         int ret = 0;
4079         struct address_space *mapping = inode->i_mapping;
4080         struct page *page;
4081         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4082                 PAGE_SHIFT;
4083
4084         struct extent_page_data epd = {
4085                 .bio = NULL,
4086                 .tree = tree,
4087                 .extent_locked = 1,
4088                 .sync_io = mode == WB_SYNC_ALL,
4089         };
4090         struct writeback_control wbc_writepages = {
4091                 .sync_mode      = mode,
4092                 .nr_to_write    = nr_pages * 2,
4093                 .range_start    = start,
4094                 .range_end      = end + 1,
4095         };
4096
4097         while (start <= end) {
4098                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4099                 if (clear_page_dirty_for_io(page))
4100                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4101                 else {
4102                         if (tree->ops && tree->ops->writepage_end_io_hook)
4103                                 tree->ops->writepage_end_io_hook(page, start,
4104                                                  start + PAGE_SIZE - 1,
4105                                                  NULL, 1);
4106                         unlock_page(page);
4107                 }
4108                 put_page(page);
4109                 start += PAGE_SIZE;
4110         }
4111
4112         flush_epd_write_bio(&epd);
4113         return ret;
4114 }
4115
4116 int extent_writepages(struct extent_io_tree *tree,
4117                       struct address_space *mapping,
4118                       struct writeback_control *wbc)
4119 {
4120         int ret = 0;
4121         struct extent_page_data epd = {
4122                 .bio = NULL,
4123                 .tree = tree,
4124                 .extent_locked = 0,
4125                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4126         };
4127
4128         ret = extent_write_cache_pages(mapping, wbc, __extent_writepage, &epd,
4129                                        flush_write_bio);
4130         flush_epd_write_bio(&epd);
4131         return ret;
4132 }
4133
4134 int extent_readpages(struct extent_io_tree *tree,
4135                      struct address_space *mapping,
4136                      struct list_head *pages, unsigned nr_pages)
4137 {
4138         struct bio *bio = NULL;
4139         unsigned page_idx;
4140         unsigned long bio_flags = 0;
4141         struct page *pagepool[16];
4142         struct page *page;
4143         struct extent_map *em_cached = NULL;
4144         int nr = 0;
4145         u64 prev_em_start = (u64)-1;
4146
4147         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4148                 page = list_entry(pages->prev, struct page, lru);
4149
4150                 prefetchw(&page->flags);
4151                 list_del(&page->lru);
4152                 if (add_to_page_cache_lru(page, mapping,
4153                                         page->index,
4154                                         readahead_gfp_mask(mapping))) {
4155                         put_page(page);
4156                         continue;
4157                 }
4158
4159                 pagepool[nr++] = page;
4160                 if (nr < ARRAY_SIZE(pagepool))
4161                         continue;
4162                 __extent_readpages(tree, pagepool, nr, &em_cached, &bio,
4163                                 &bio_flags, &prev_em_start);
4164                 nr = 0;
4165         }
4166         if (nr)
4167                 __extent_readpages(tree, pagepool, nr, &em_cached, &bio,
4168                                 &bio_flags, &prev_em_start);
4169
4170         if (em_cached)
4171                 free_extent_map(em_cached);
4172
4173         BUG_ON(!list_empty(pages));
4174         if (bio)
4175                 return submit_one_bio(bio, 0, bio_flags);
4176         return 0;
4177 }
4178
4179 /*
4180  * basic invalidatepage code, this waits on any locked or writeback
4181  * ranges corresponding to the page, and then deletes any extent state
4182  * records from the tree
4183  */
4184 int extent_invalidatepage(struct extent_io_tree *tree,
4185                           struct page *page, unsigned long offset)
4186 {
4187         struct extent_state *cached_state = NULL;
4188         u64 start = page_offset(page);
4189         u64 end = start + PAGE_SIZE - 1;
4190         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4191
4192         start += ALIGN(offset, blocksize);
4193         if (start > end)
4194                 return 0;
4195
4196         lock_extent_bits(tree, start, end, &cached_state);
4197         wait_on_page_writeback(page);
4198         clear_extent_bit(tree, start, end,
4199                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4200                          EXTENT_DO_ACCOUNTING,
4201                          1, 1, &cached_state);
4202         return 0;
4203 }
4204
4205 /*
4206  * a helper for releasepage, this tests for areas of the page that
4207  * are locked or under IO and drops the related state bits if it is safe
4208  * to drop the page.
4209  */
4210 static int try_release_extent_state(struct extent_map_tree *map,
4211                                     struct extent_io_tree *tree,
4212                                     struct page *page, gfp_t mask)
4213 {
4214         u64 start = page_offset(page);
4215         u64 end = start + PAGE_SIZE - 1;
4216         int ret = 1;
4217
4218         if (test_range_bit(tree, start, end,
4219                            EXTENT_IOBITS, 0, NULL))
4220                 ret = 0;
4221         else {
4222                 /*
4223                  * at this point we can safely clear everything except the
4224                  * locked bit and the nodatasum bit
4225                  */
4226                 ret = __clear_extent_bit(tree, start, end,
4227                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4228                                  0, 0, NULL, mask, NULL);
4229
4230                 /* if clear_extent_bit failed for enomem reasons,
4231                  * we can't allow the release to continue.
4232                  */
4233                 if (ret < 0)
4234                         ret = 0;
4235                 else
4236                         ret = 1;
4237         }
4238         return ret;
4239 }
4240
4241 /*
4242  * a helper for releasepage.  As long as there are no locked extents
4243  * in the range corresponding to the page, both state records and extent
4244  * map records are removed
4245  */
4246 int try_release_extent_mapping(struct extent_map_tree *map,
4247                                struct extent_io_tree *tree, struct page *page,
4248                                gfp_t mask)
4249 {
4250         struct extent_map *em;
4251         u64 start = page_offset(page);
4252         u64 end = start + PAGE_SIZE - 1;
4253
4254         if (gfpflags_allow_blocking(mask) &&
4255             page->mapping->host->i_size > SZ_16M) {
4256                 u64 len;
4257                 while (start <= end) {
4258                         len = end - start + 1;
4259                         write_lock(&map->lock);
4260                         em = lookup_extent_mapping(map, start, len);
4261                         if (!em) {
4262                                 write_unlock(&map->lock);
4263                                 break;
4264                         }
4265                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4266                             em->start != start) {
4267                                 write_unlock(&map->lock);
4268                                 free_extent_map(em);
4269                                 break;
4270                         }
4271                         if (!test_range_bit(tree, em->start,
4272                                             extent_map_end(em) - 1,
4273                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4274                                             0, NULL)) {
4275                                 remove_extent_mapping(map, em);
4276                                 /* once for the rb tree */
4277                                 free_extent_map(em);
4278                         }
4279                         start = extent_map_end(em);
4280                         write_unlock(&map->lock);
4281
4282                         /* once for us */
4283                         free_extent_map(em);
4284                 }
4285         }
4286         return try_release_extent_state(map, tree, page, mask);
4287 }
4288
4289 /*
4290  * helper function for fiemap, which doesn't want to see any holes.
4291  * This maps until we find something past 'last'
4292  */
4293 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4294                                                 u64 offset, u64 last)
4295 {
4296         u64 sectorsize = btrfs_inode_sectorsize(inode);
4297         struct extent_map *em;
4298         u64 len;
4299
4300         if (offset >= last)
4301                 return NULL;
4302
4303         while (1) {
4304                 len = last - offset;
4305                 if (len == 0)
4306                         break;
4307                 len = ALIGN(len, sectorsize);
4308                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), NULL, 0, offset,
4309                                 len, 0);
4310                 if (IS_ERR_OR_NULL(em))
4311                         return em;
4312
4313                 /* if this isn't a hole return it */
4314                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4315                     em->block_start != EXTENT_MAP_HOLE) {
4316                         return em;
4317                 }
4318
4319                 /* this is a hole, advance to the next extent */
4320                 offset = extent_map_end(em);
4321                 free_extent_map(em);
4322                 if (offset >= last)
4323                         break;
4324         }
4325         return NULL;
4326 }
4327
4328 /*
4329  * To cache previous fiemap extent
4330  *
4331  * Will be used for merging fiemap extent
4332  */
4333 struct fiemap_cache {
4334         u64 offset;
4335         u64 phys;
4336         u64 len;
4337         u32 flags;
4338         bool cached;
4339 };
4340
4341 /*
4342  * Helper to submit fiemap extent.
4343  *
4344  * Will try to merge current fiemap extent specified by @offset, @phys,
4345  * @len and @flags with cached one.
4346  * And only when we fails to merge, cached one will be submitted as
4347  * fiemap extent.
4348  *
4349  * Return value is the same as fiemap_fill_next_extent().
4350  */
4351 static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
4352                                 struct fiemap_cache *cache,
4353                                 u64 offset, u64 phys, u64 len, u32 flags)
4354 {
4355         int ret = 0;
4356
4357         if (!cache->cached)
4358                 goto assign;
4359
4360         /*
4361          * Sanity check, extent_fiemap() should have ensured that new
4362          * fiemap extent won't overlap with cahced one.
4363          * Not recoverable.
4364          *
4365          * NOTE: Physical address can overlap, due to compression
4366          */
4367         if (cache->offset + cache->len > offset) {
4368                 WARN_ON(1);
4369                 return -EINVAL;
4370         }
4371
4372         /*
4373          * Only merges fiemap extents if
4374          * 1) Their logical addresses are continuous
4375          *
4376          * 2) Their physical addresses are continuous
4377          *    So truly compressed (physical size smaller than logical size)
4378          *    extents won't get merged with each other
4379          *
4380          * 3) Share same flags except FIEMAP_EXTENT_LAST
4381          *    So regular extent won't get merged with prealloc extent
4382          */
4383         if (cache->offset + cache->len  == offset &&
4384             cache->phys + cache->len == phys  &&
4385             (cache->flags & ~FIEMAP_EXTENT_LAST) ==
4386                         (flags & ~FIEMAP_EXTENT_LAST)) {
4387                 cache->len += len;
4388                 cache->flags |= flags;
4389                 goto try_submit_last;
4390         }
4391
4392         /* Not mergeable, need to submit cached one */
4393         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4394                                       cache->len, cache->flags);
4395         cache->cached = false;
4396         if (ret)
4397                 return ret;
4398 assign:
4399         cache->cached = true;
4400         cache->offset = offset;
4401         cache->phys = phys;
4402         cache->len = len;
4403         cache->flags = flags;
4404 try_submit_last:
4405         if (cache->flags & FIEMAP_EXTENT_LAST) {
4406                 ret = fiemap_fill_next_extent(fieinfo, cache->offset,
4407                                 cache->phys, cache->len, cache->flags);
4408                 cache->cached = false;
4409         }
4410         return ret;
4411 }
4412
4413 /*
4414  * Emit last fiemap cache
4415  *
4416  * The last fiemap cache may still be cached in the following case:
4417  * 0                  4k                    8k
4418  * |<- Fiemap range ->|
4419  * |<------------  First extent ----------->|
4420  *
4421  * In this case, the first extent range will be cached but not emitted.
4422  * So we must emit it before ending extent_fiemap().
4423  */
4424 static int emit_last_fiemap_cache(struct btrfs_fs_info *fs_info,
4425                                   struct fiemap_extent_info *fieinfo,
4426                                   struct fiemap_cache *cache)
4427 {
4428         int ret;
4429
4430         if (!cache->cached)
4431                 return 0;
4432
4433         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4434                                       cache->len, cache->flags);
4435         cache->cached = false;
4436         if (ret > 0)
4437                 ret = 0;
4438         return ret;
4439 }
4440
4441 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4442                 __u64 start, __u64 len)
4443 {
4444         int ret = 0;
4445         u64 off = start;
4446         u64 max = start + len;
4447         u32 flags = 0;
4448         u32 found_type;
4449         u64 last;
4450         u64 last_for_get_extent = 0;
4451         u64 disko = 0;
4452         u64 isize = i_size_read(inode);
4453         struct btrfs_key found_key;
4454         struct extent_map *em = NULL;
4455         struct extent_state *cached_state = NULL;
4456         struct btrfs_path *path;
4457         struct btrfs_root *root = BTRFS_I(inode)->root;
4458         struct fiemap_cache cache = { 0 };
4459         int end = 0;
4460         u64 em_start = 0;
4461         u64 em_len = 0;
4462         u64 em_end = 0;
4463
4464         if (len == 0)
4465                 return -EINVAL;
4466
4467         path = btrfs_alloc_path();
4468         if (!path)
4469                 return -ENOMEM;
4470         path->leave_spinning = 1;
4471
4472         start = round_down(start, btrfs_inode_sectorsize(inode));
4473         len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
4474
4475         /*
4476          * lookup the last file extent.  We're not using i_size here
4477          * because there might be preallocation past i_size
4478          */
4479         ret = btrfs_lookup_file_extent(NULL, root, path,
4480                         btrfs_ino(BTRFS_I(inode)), -1, 0);
4481         if (ret < 0) {
4482                 btrfs_free_path(path);
4483                 return ret;
4484         } else {
4485                 WARN_ON(!ret);
4486                 if (ret == 1)
4487                         ret = 0;
4488         }
4489
4490         path->slots[0]--;
4491         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4492         found_type = found_key.type;
4493
4494         /* No extents, but there might be delalloc bits */
4495         if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) ||
4496             found_type != BTRFS_EXTENT_DATA_KEY) {
4497                 /* have to trust i_size as the end */
4498                 last = (u64)-1;
4499                 last_for_get_extent = isize;
4500         } else {
4501                 /*
4502                  * remember the start of the last extent.  There are a
4503                  * bunch of different factors that go into the length of the
4504                  * extent, so its much less complex to remember where it started
4505                  */
4506                 last = found_key.offset;
4507                 last_for_get_extent = last + 1;
4508         }
4509         btrfs_release_path(path);
4510
4511         /*
4512          * we might have some extents allocated but more delalloc past those
4513          * extents.  so, we trust isize unless the start of the last extent is
4514          * beyond isize
4515          */
4516         if (last < isize) {
4517                 last = (u64)-1;
4518                 last_for_get_extent = isize;
4519         }
4520
4521         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4522                          &cached_state);
4523
4524         em = get_extent_skip_holes(inode, start, last_for_get_extent);
4525         if (!em)
4526                 goto out;
4527         if (IS_ERR(em)) {
4528                 ret = PTR_ERR(em);
4529                 goto out;
4530         }
4531
4532         while (!end) {
4533                 u64 offset_in_extent = 0;
4534
4535                 /* break if the extent we found is outside the range */
4536                 if (em->start >= max || extent_map_end(em) < off)
4537                         break;
4538
4539                 /*
4540                  * get_extent may return an extent that starts before our
4541                  * requested range.  We have to make sure the ranges
4542                  * we return to fiemap always move forward and don't
4543                  * overlap, so adjust the offsets here
4544                  */
4545                 em_start = max(em->start, off);
4546
4547                 /*
4548                  * record the offset from the start of the extent
4549                  * for adjusting the disk offset below.  Only do this if the
4550                  * extent isn't compressed since our in ram offset may be past
4551                  * what we have actually allocated on disk.
4552                  */
4553                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4554                         offset_in_extent = em_start - em->start;
4555                 em_end = extent_map_end(em);
4556                 em_len = em_end - em_start;
4557                 disko = 0;
4558                 flags = 0;
4559
4560                 /*
4561                  * bump off for our next call to get_extent
4562                  */
4563                 off = extent_map_end(em);
4564                 if (off >= max)
4565                         end = 1;
4566
4567                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4568                         end = 1;
4569                         flags |= FIEMAP_EXTENT_LAST;
4570                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4571                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4572                                   FIEMAP_EXTENT_NOT_ALIGNED);
4573                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4574                         flags |= (FIEMAP_EXTENT_DELALLOC |
4575                                   FIEMAP_EXTENT_UNKNOWN);
4576                 } else if (fieinfo->fi_extents_max) {
4577                         u64 bytenr = em->block_start -
4578                                 (em->start - em->orig_start);
4579
4580                         disko = em->block_start + offset_in_extent;
4581
4582                         /*
4583                          * As btrfs supports shared space, this information
4584                          * can be exported to userspace tools via
4585                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4586                          * then we're just getting a count and we can skip the
4587                          * lookup stuff.
4588                          */
4589                         ret = btrfs_check_shared(root,
4590                                                  btrfs_ino(BTRFS_I(inode)),
4591                                                  bytenr);
4592                         if (ret < 0)
4593                                 goto out_free;
4594                         if (ret)
4595                                 flags |= FIEMAP_EXTENT_SHARED;
4596                         ret = 0;
4597                 }
4598                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4599                         flags |= FIEMAP_EXTENT_ENCODED;
4600                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4601                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4602
4603                 free_extent_map(em);
4604                 em = NULL;
4605                 if ((em_start >= last) || em_len == (u64)-1 ||
4606                    (last == (u64)-1 && isize <= em_end)) {
4607                         flags |= FIEMAP_EXTENT_LAST;
4608                         end = 1;
4609                 }
4610
4611                 /* now scan forward to see if this is really the last extent. */
4612                 em = get_extent_skip_holes(inode, off, last_for_get_extent);
4613                 if (IS_ERR(em)) {
4614                         ret = PTR_ERR(em);
4615                         goto out;
4616                 }
4617                 if (!em) {
4618                         flags |= FIEMAP_EXTENT_LAST;
4619                         end = 1;
4620                 }
4621                 ret = emit_fiemap_extent(fieinfo, &cache, em_start, disko,
4622                                            em_len, flags);
4623                 if (ret) {
4624                         if (ret == 1)
4625                                 ret = 0;
4626                         goto out_free;
4627                 }
4628         }
4629 out_free:
4630         if (!ret)
4631                 ret = emit_last_fiemap_cache(root->fs_info, fieinfo, &cache);
4632         free_extent_map(em);
4633 out:
4634         btrfs_free_path(path);
4635         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4636                              &cached_state, GFP_NOFS);
4637         return ret;
4638 }
4639
4640 static void __free_extent_buffer(struct extent_buffer *eb)
4641 {
4642         btrfs_leak_debug_del(&eb->leak_list);
4643         kmem_cache_free(extent_buffer_cache, eb);
4644 }
4645
4646 int extent_buffer_under_io(struct extent_buffer *eb)
4647 {
4648         return (atomic_read(&eb->io_pages) ||
4649                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4650                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4651 }
4652
4653 /*
4654  * Helper for releasing extent buffer page.
4655  */
4656 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4657 {
4658         unsigned long index;
4659         struct page *page;
4660         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4661
4662         BUG_ON(extent_buffer_under_io(eb));
4663
4664         index = num_extent_pages(eb->start, eb->len);
4665         if (index == 0)
4666                 return;
4667
4668         do {
4669                 index--;
4670                 page = eb->pages[index];
4671                 if (!page)
4672                         continue;
4673                 if (mapped)
4674                         spin_lock(&page->mapping->private_lock);
4675                 /*
4676                  * We do this since we'll remove the pages after we've
4677                  * removed the eb from the radix tree, so we could race
4678                  * and have this page now attached to the new eb.  So
4679                  * only clear page_private if it's still connected to
4680                  * this eb.
4681                  */
4682                 if (PagePrivate(page) &&
4683                     page->private == (unsigned long)eb) {
4684                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4685                         BUG_ON(PageDirty(page));
4686                         BUG_ON(PageWriteback(page));
4687                         /*
4688                          * We need to make sure we haven't be attached
4689                          * to a new eb.
4690                          */
4691                         ClearPagePrivate(page);
4692                         set_page_private(page, 0);
4693                         /* One for the page private */
4694                         put_page(page);
4695                 }
4696
4697                 if (mapped)
4698                         spin_unlock(&page->mapping->private_lock);
4699
4700                 /* One for when we allocated the page */
4701                 put_page(page);
4702         } while (index != 0);
4703 }
4704
4705 /*
4706  * Helper for releasing the extent buffer.
4707  */
4708 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4709 {
4710         btrfs_release_extent_buffer_page(eb);
4711         __free_extent_buffer(eb);
4712 }
4713
4714 static struct extent_buffer *
4715 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4716                       unsigned long len)
4717 {
4718         struct extent_buffer *eb = NULL;
4719
4720         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4721         eb->start = start;
4722         eb->len = len;
4723         eb->fs_info = fs_info;
4724         eb->bflags = 0;
4725         rwlock_init(&eb->lock);
4726         atomic_set(&eb->write_locks, 0);
4727         atomic_set(&eb->read_locks, 0);
4728         atomic_set(&eb->blocking_readers, 0);
4729         atomic_set(&eb->blocking_writers, 0);
4730         atomic_set(&eb->spinning_readers, 0);
4731         atomic_set(&eb->spinning_writers, 0);
4732         eb->lock_nested = 0;
4733         init_waitqueue_head(&eb->write_lock_wq);
4734         init_waitqueue_head(&eb->read_lock_wq);
4735
4736         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4737
4738         spin_lock_init(&eb->refs_lock);
4739         atomic_set(&eb->refs, 1);
4740         atomic_set(&eb->io_pages, 0);
4741
4742         /*
4743          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4744          */
4745         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4746                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4747         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4748
4749         return eb;
4750 }
4751
4752 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4753 {
4754         unsigned long i;
4755         struct page *p;
4756         struct extent_buffer *new;
4757         unsigned long num_pages = num_extent_pages(src->start, src->len);
4758
4759         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4760         if (new == NULL)
4761                 return NULL;
4762
4763         for (i = 0; i < num_pages; i++) {
4764                 p = alloc_page(GFP_NOFS);
4765                 if (!p) {
4766                         btrfs_release_extent_buffer(new);
4767                         return NULL;
4768                 }
4769                 attach_extent_buffer_page(new, p);
4770                 WARN_ON(PageDirty(p));
4771                 SetPageUptodate(p);
4772                 new->pages[i] = p;
4773                 copy_page(page_address(p), page_address(src->pages[i]));
4774         }
4775
4776         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4777         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4778
4779         return new;
4780 }
4781
4782 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4783                                                   u64 start, unsigned long len)
4784 {
4785         struct extent_buffer *eb;
4786         unsigned long num_pages;
4787         unsigned long i;
4788
4789         num_pages = num_extent_pages(start, len);
4790
4791         eb = __alloc_extent_buffer(fs_info, start, len);
4792         if (!eb)
4793                 return NULL;
4794
4795         for (i = 0; i < num_pages; i++) {
4796                 eb->pages[i] = alloc_page(GFP_NOFS);
4797                 if (!eb->pages[i])
4798                         goto err;
4799         }
4800         set_extent_buffer_uptodate(eb);
4801         btrfs_set_header_nritems(eb, 0);
4802         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4803
4804         return eb;
4805 err:
4806         for (; i > 0; i--)
4807                 __free_page(eb->pages[i - 1]);
4808         __free_extent_buffer(eb);
4809         return NULL;
4810 }
4811
4812 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4813                                                 u64 start)
4814 {
4815         return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
4816 }
4817
4818 static void check_buffer_tree_ref(struct extent_buffer *eb)
4819 {
4820         int refs;
4821         /* the ref bit is tricky.  We have to make sure it is set
4822          * if we have the buffer dirty.   Otherwise the
4823          * code to free a buffer can end up dropping a dirty
4824          * page
4825          *
4826          * Once the ref bit is set, it won't go away while the
4827          * buffer is dirty or in writeback, and it also won't
4828          * go away while we have the reference count on the
4829          * eb bumped.
4830          *
4831          * We can't just set the ref bit without bumping the
4832          * ref on the eb because free_extent_buffer might
4833          * see the ref bit and try to clear it.  If this happens
4834          * free_extent_buffer might end up dropping our original
4835          * ref by mistake and freeing the page before we are able
4836          * to add one more ref.
4837          *
4838          * So bump the ref count first, then set the bit.  If someone
4839          * beat us to it, drop the ref we added.
4840          */
4841         refs = atomic_read(&eb->refs);
4842         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4843                 return;
4844
4845         spin_lock(&eb->refs_lock);
4846         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4847                 atomic_inc(&eb->refs);
4848         spin_unlock(&eb->refs_lock);
4849 }
4850
4851 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4852                 struct page *accessed)
4853 {
4854         unsigned long num_pages, i;
4855
4856         check_buffer_tree_ref(eb);
4857
4858         num_pages = num_extent_pages(eb->start, eb->len);
4859         for (i = 0; i < num_pages; i++) {
4860                 struct page *p = eb->pages[i];
4861
4862                 if (p != accessed)
4863                         mark_page_accessed(p);
4864         }
4865 }
4866
4867 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4868                                          u64 start)
4869 {
4870         struct extent_buffer *eb;
4871
4872         rcu_read_lock();
4873         eb = radix_tree_lookup(&fs_info->buffer_radix,
4874                                start >> PAGE_SHIFT);
4875         if (eb && atomic_inc_not_zero(&eb->refs)) {
4876                 rcu_read_unlock();
4877                 /*
4878                  * Lock our eb's refs_lock to avoid races with
4879                  * free_extent_buffer. When we get our eb it might be flagged
4880                  * with EXTENT_BUFFER_STALE and another task running
4881                  * free_extent_buffer might have seen that flag set,
4882                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4883                  * writeback flags not set) and it's still in the tree (flag
4884                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4885                  * of decrementing the extent buffer's reference count twice.
4886                  * So here we could race and increment the eb's reference count,
4887                  * clear its stale flag, mark it as dirty and drop our reference
4888                  * before the other task finishes executing free_extent_buffer,
4889                  * which would later result in an attempt to free an extent
4890                  * buffer that is dirty.
4891                  */
4892                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4893                         spin_lock(&eb->refs_lock);
4894                         spin_unlock(&eb->refs_lock);
4895                 }
4896                 mark_extent_buffer_accessed(eb, NULL);
4897                 return eb;
4898         }
4899         rcu_read_unlock();
4900
4901         return NULL;
4902 }
4903
4904 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4905 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4906                                         u64 start)
4907 {
4908         struct extent_buffer *eb, *exists = NULL;
4909         int ret;
4910
4911         eb = find_extent_buffer(fs_info, start);
4912         if (eb)
4913                 return eb;
4914         eb = alloc_dummy_extent_buffer(fs_info, start);
4915         if (!eb)
4916                 return NULL;
4917         eb->fs_info = fs_info;
4918 again:
4919         ret = radix_tree_preload(GFP_NOFS);
4920         if (ret)
4921                 goto free_eb;
4922         spin_lock(&fs_info->buffer_lock);
4923         ret = radix_tree_insert(&fs_info->buffer_radix,
4924                                 start >> PAGE_SHIFT, eb);
4925         spin_unlock(&fs_info->buffer_lock);
4926         radix_tree_preload_end();
4927         if (ret == -EEXIST) {
4928                 exists = find_extent_buffer(fs_info, start);
4929                 if (exists)
4930                         goto free_eb;
4931                 else
4932                         goto again;
4933         }
4934         check_buffer_tree_ref(eb);
4935         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4936
4937         /*
4938          * We will free dummy extent buffer's if they come into
4939          * free_extent_buffer with a ref count of 2, but if we are using this we
4940          * want the buffers to stay in memory until we're done with them, so
4941          * bump the ref count again.
4942          */
4943         atomic_inc(&eb->refs);
4944         return eb;
4945 free_eb:
4946         btrfs_release_extent_buffer(eb);
4947         return exists;
4948 }
4949 #endif
4950
4951 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4952                                           u64 start)
4953 {
4954         unsigned long len = fs_info->nodesize;
4955         unsigned long num_pages = num_extent_pages(start, len);
4956         unsigned long i;
4957         unsigned long index = start >> PAGE_SHIFT;
4958         struct extent_buffer *eb;
4959         struct extent_buffer *exists = NULL;
4960         struct page *p;
4961         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4962         int uptodate = 1;
4963         int ret;
4964
4965         if (!IS_ALIGNED(start, fs_info->sectorsize)) {
4966                 btrfs_err(fs_info, "bad tree block start %llu", start);
4967                 return ERR_PTR(-EINVAL);
4968         }
4969
4970         eb = find_extent_buffer(fs_info, start);
4971         if (eb)
4972                 return eb;
4973
4974         eb = __alloc_extent_buffer(fs_info, start, len);
4975         if (!eb)
4976                 return ERR_PTR(-ENOMEM);
4977
4978         for (i = 0; i < num_pages; i++, index++) {
4979                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4980                 if (!p) {
4981                         exists = ERR_PTR(-ENOMEM);
4982                         goto free_eb;
4983                 }
4984
4985                 spin_lock(&mapping->private_lock);
4986                 if (PagePrivate(p)) {
4987                         /*
4988                          * We could have already allocated an eb for this page
4989                          * and attached one so lets see if we can get a ref on
4990                          * the existing eb, and if we can we know it's good and
4991                          * we can just return that one, else we know we can just
4992                          * overwrite page->private.
4993                          */
4994                         exists = (struct extent_buffer *)p->private;
4995                         if (atomic_inc_not_zero(&exists->refs)) {
4996                                 spin_unlock(&mapping->private_lock);
4997                                 unlock_page(p);
4998                                 put_page(p);
4999                                 mark_extent_buffer_accessed(exists, p);
5000                                 goto free_eb;
5001                         }
5002                         exists = NULL;
5003
5004                         /*
5005                          * Do this so attach doesn't complain and we need to
5006                          * drop the ref the old guy had.
5007                          */
5008                         ClearPagePrivate(p);
5009                         WARN_ON(PageDirty(p));
5010                         put_page(p);
5011                 }
5012                 attach_extent_buffer_page(eb, p);
5013                 spin_unlock(&mapping->private_lock);
5014                 WARN_ON(PageDirty(p));
5015                 eb->pages[i] = p;
5016                 if (!PageUptodate(p))
5017                         uptodate = 0;
5018
5019                 /*
5020                  * see below about how we avoid a nasty race with release page
5021                  * and why we unlock later
5022                  */
5023         }
5024         if (uptodate)
5025                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5026 again:
5027         ret = radix_tree_preload(GFP_NOFS);
5028         if (ret) {
5029                 exists = ERR_PTR(ret);
5030                 goto free_eb;
5031         }
5032
5033         spin_lock(&fs_info->buffer_lock);
5034         ret = radix_tree_insert(&fs_info->buffer_radix,
5035                                 start >> PAGE_SHIFT, eb);
5036         spin_unlock(&fs_info->buffer_lock);
5037         radix_tree_preload_end();
5038         if (ret == -EEXIST) {
5039                 exists = find_extent_buffer(fs_info, start);
5040                 if (exists)
5041                         goto free_eb;
5042                 else
5043                         goto again;
5044         }
5045         /* add one reference for the tree */
5046         check_buffer_tree_ref(eb);
5047         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5048
5049         /*
5050          * there is a race where release page may have
5051          * tried to find this extent buffer in the radix
5052          * but failed.  It will tell the VM it is safe to
5053          * reclaim the, and it will clear the page private bit.
5054          * We must make sure to set the page private bit properly
5055          * after the extent buffer is in the radix tree so
5056          * it doesn't get lost
5057          */
5058         SetPageChecked(eb->pages[0]);
5059         for (i = 1; i < num_pages; i++) {
5060                 p = eb->pages[i];
5061                 ClearPageChecked(p);
5062                 unlock_page(p);
5063         }
5064         unlock_page(eb->pages[0]);
5065         return eb;
5066
5067 free_eb:
5068         WARN_ON(!atomic_dec_and_test(&eb->refs));
5069         for (i = 0; i < num_pages; i++) {
5070                 if (eb->pages[i])
5071                         unlock_page(eb->pages[i]);
5072         }
5073
5074         btrfs_release_extent_buffer(eb);
5075         return exists;
5076 }
5077
5078 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5079 {
5080         struct extent_buffer *eb =
5081                         container_of(head, struct extent_buffer, rcu_head);
5082
5083         __free_extent_buffer(eb);
5084 }
5085
5086 /* Expects to have eb->eb_lock already held */
5087 static int release_extent_buffer(struct extent_buffer *eb)
5088 {
5089         WARN_ON(atomic_read(&eb->refs) == 0);
5090         if (atomic_dec_and_test(&eb->refs)) {
5091                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5092                         struct btrfs_fs_info *fs_info = eb->fs_info;
5093
5094                         spin_unlock(&eb->refs_lock);
5095
5096                         spin_lock(&fs_info->buffer_lock);
5097                         radix_tree_delete(&fs_info->buffer_radix,
5098                                           eb->start >> PAGE_SHIFT);
5099                         spin_unlock(&fs_info->buffer_lock);
5100                 } else {
5101                         spin_unlock(&eb->refs_lock);
5102                 }
5103
5104                 /* Should be safe to release our pages at this point */
5105                 btrfs_release_extent_buffer_page(eb);
5106 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5107                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5108                         __free_extent_buffer(eb);
5109                         return 1;
5110                 }
5111 #endif
5112                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5113                 return 1;
5114         }
5115         spin_unlock(&eb->refs_lock);
5116
5117         return 0;
5118 }
5119
5120 void free_extent_buffer(struct extent_buffer *eb)
5121 {
5122         int refs;
5123         int old;
5124         if (!eb)
5125                 return;
5126
5127         while (1) {
5128                 refs = atomic_read(&eb->refs);
5129                 if (refs <= 3)
5130                         break;
5131                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5132                 if (old == refs)
5133                         return;
5134         }
5135
5136         spin_lock(&eb->refs_lock);
5137         if (atomic_read(&eb->refs) == 2 &&
5138             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5139                 atomic_dec(&eb->refs);
5140
5141         if (atomic_read(&eb->refs) == 2 &&
5142             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5143             !extent_buffer_under_io(eb) &&
5144             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5145                 atomic_dec(&eb->refs);
5146
5147         /*
5148          * I know this is terrible, but it's temporary until we stop tracking
5149          * the uptodate bits and such for the extent buffers.
5150          */
5151         release_extent_buffer(eb);
5152 }
5153
5154 void free_extent_buffer_stale(struct extent_buffer *eb)
5155 {
5156         if (!eb)
5157                 return;
5158
5159         spin_lock(&eb->refs_lock);
5160         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5161
5162         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5163             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5164                 atomic_dec(&eb->refs);
5165         release_extent_buffer(eb);
5166 }
5167
5168 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5169 {
5170         unsigned long i;
5171         unsigned long num_pages;
5172         struct page *page;
5173
5174         num_pages = num_extent_pages(eb->start, eb->len);
5175
5176         for (i = 0; i < num_pages; i++) {
5177                 page = eb->pages[i];
5178                 if (!PageDirty(page))
5179                         continue;
5180
5181                 lock_page(page);
5182                 WARN_ON(!PagePrivate(page));
5183
5184                 clear_page_dirty_for_io(page);
5185                 spin_lock_irq(&page->mapping->tree_lock);
5186                 if (!PageDirty(page)) {
5187                         radix_tree_tag_clear(&page->mapping->page_tree,
5188                                                 page_index(page),
5189                                                 PAGECACHE_TAG_DIRTY);
5190                 }
5191                 spin_unlock_irq(&page->mapping->tree_lock);
5192                 ClearPageError(page);
5193                 unlock_page(page);
5194         }
5195         WARN_ON(atomic_read(&eb->refs) == 0);
5196 }
5197
5198 int set_extent_buffer_dirty(struct extent_buffer *eb)
5199 {
5200         unsigned long i;
5201         unsigned long num_pages;
5202         int was_dirty = 0;
5203
5204         check_buffer_tree_ref(eb);
5205
5206         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5207
5208         num_pages = num_extent_pages(eb->start, eb->len);
5209         WARN_ON(atomic_read(&eb->refs) == 0);
5210         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5211
5212         for (i = 0; i < num_pages; i++)
5213                 set_page_dirty(eb->pages[i]);
5214         return was_dirty;
5215 }
5216
5217 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5218 {
5219         unsigned long i;
5220         struct page *page;
5221         unsigned long num_pages;
5222
5223         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5224         num_pages = num_extent_pages(eb->start, eb->len);
5225         for (i = 0; i < num_pages; i++) {
5226                 page = eb->pages[i];
5227                 if (page)
5228                         ClearPageUptodate(page);
5229         }
5230 }
5231
5232 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5233 {
5234         unsigned long i;
5235         struct page *page;
5236         unsigned long num_pages;
5237
5238         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5239         num_pages = num_extent_pages(eb->start, eb->len);
5240         for (i = 0; i < num_pages; i++) {
5241                 page = eb->pages[i];
5242                 SetPageUptodate(page);
5243         }
5244 }
5245
5246 int extent_buffer_uptodate(struct extent_buffer *eb)
5247 {
5248         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5249 }
5250
5251 int read_extent_buffer_pages(struct extent_io_tree *tree,
5252                              struct extent_buffer *eb, int wait, int mirror_num)
5253 {
5254         unsigned long i;
5255         struct page *page;
5256         int err;
5257         int ret = 0;
5258         int locked_pages = 0;
5259         int all_uptodate = 1;
5260         unsigned long num_pages;
5261         unsigned long num_reads = 0;
5262         struct bio *bio = NULL;
5263         unsigned long bio_flags = 0;
5264
5265         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5266                 return 0;
5267
5268         num_pages = num_extent_pages(eb->start, eb->len);
5269         for (i = 0; i < num_pages; i++) {
5270                 page = eb->pages[i];
5271                 if (wait == WAIT_NONE) {
5272                         if (!trylock_page(page))
5273                                 goto unlock_exit;
5274                 } else {
5275                         lock_page(page);
5276                 }
5277                 locked_pages++;
5278         }
5279         /*
5280          * We need to firstly lock all pages to make sure that
5281          * the uptodate bit of our pages won't be affected by
5282          * clear_extent_buffer_uptodate().
5283          */
5284         for (i = 0; i < num_pages; i++) {
5285                 page = eb->pages[i];
5286                 if (!PageUptodate(page)) {
5287                         num_reads++;
5288                         all_uptodate = 0;
5289                 }
5290         }
5291
5292         if (all_uptodate) {
5293                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5294                 goto unlock_exit;
5295         }
5296
5297         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5298         eb->read_mirror = 0;
5299         atomic_set(&eb->io_pages, num_reads);
5300         for (i = 0; i < num_pages; i++) {
5301                 page = eb->pages[i];
5302
5303                 if (!PageUptodate(page)) {
5304                         if (ret) {
5305                                 atomic_dec(&eb->io_pages);
5306                                 unlock_page(page);
5307                                 continue;
5308                         }
5309
5310                         ClearPageError(page);
5311                         err = __extent_read_full_page(tree, page,
5312                                                       btree_get_extent, &bio,
5313                                                       mirror_num, &bio_flags,
5314                                                       REQ_META);
5315                         if (err) {
5316                                 ret = err;
5317                                 /*
5318                                  * We use &bio in above __extent_read_full_page,
5319                                  * so we ensure that if it returns error, the
5320                                  * current page fails to add itself to bio and
5321                                  * it's been unlocked.
5322                                  *
5323                                  * We must dec io_pages by ourselves.
5324                                  */
5325                                 atomic_dec(&eb->io_pages);
5326                         }
5327                 } else {
5328                         unlock_page(page);
5329                 }
5330         }
5331
5332         if (bio) {
5333                 err = submit_one_bio(bio, mirror_num, bio_flags);
5334                 if (err)
5335                         return err;
5336         }
5337
5338         if (ret || wait != WAIT_COMPLETE)
5339                 return ret;
5340
5341         for (i = 0; i < num_pages; i++) {
5342                 page = eb->pages[i];
5343                 wait_on_page_locked(page);
5344                 if (!PageUptodate(page))
5345                         ret = -EIO;
5346         }
5347
5348         return ret;
5349
5350 unlock_exit:
5351         while (locked_pages > 0) {
5352                 locked_pages--;
5353                 page = eb->pages[locked_pages];
5354                 unlock_page(page);
5355         }
5356         return ret;
5357 }
5358
5359 void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
5360                         unsigned long start, unsigned long len)
5361 {
5362         size_t cur;
5363         size_t offset;
5364         struct page *page;
5365         char *kaddr;
5366         char *dst = (char *)dstv;
5367         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5368         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5369
5370         if (start + len > eb->len) {
5371                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5372                      eb->start, eb->len, start, len);
5373                 memset(dst, 0, len);
5374                 return;
5375         }
5376
5377         offset = (start_offset + start) & (PAGE_SIZE - 1);
5378
5379         while (len > 0) {
5380                 page = eb->pages[i];
5381
5382                 cur = min(len, (PAGE_SIZE - offset));
5383                 kaddr = page_address(page);
5384                 memcpy(dst, kaddr + offset, cur);
5385
5386                 dst += cur;
5387                 len -= cur;
5388                 offset = 0;
5389                 i++;
5390         }
5391 }
5392
5393 int read_extent_buffer_to_user(const struct extent_buffer *eb,
5394                                void __user *dstv,
5395                                unsigned long start, unsigned long len)
5396 {
5397         size_t cur;
5398         size_t offset;
5399         struct page *page;
5400         char *kaddr;
5401         char __user *dst = (char __user *)dstv;
5402         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5403         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5404         int ret = 0;
5405
5406         WARN_ON(start > eb->len);
5407         WARN_ON(start + len > eb->start + eb->len);
5408
5409         offset = (start_offset + start) & (PAGE_SIZE - 1);
5410
5411         while (len > 0) {
5412                 page = eb->pages[i];
5413
5414                 cur = min(len, (PAGE_SIZE - offset));
5415                 kaddr = page_address(page);
5416                 if (copy_to_user(dst, kaddr + offset, cur)) {
5417                         ret = -EFAULT;
5418                         break;
5419                 }
5420
5421                 dst += cur;
5422                 len -= cur;
5423                 offset = 0;
5424                 i++;
5425         }
5426
5427         return ret;
5428 }
5429
5430 /*
5431  * return 0 if the item is found within a page.
5432  * return 1 if the item spans two pages.
5433  * return -EINVAL otherwise.
5434  */
5435 int map_private_extent_buffer(const struct extent_buffer *eb,
5436                               unsigned long start, unsigned long min_len,
5437                               char **map, unsigned long *map_start,
5438                               unsigned long *map_len)
5439 {
5440         size_t offset = start & (PAGE_SIZE - 1);
5441         char *kaddr;
5442         struct page *p;
5443         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5444         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5445         unsigned long end_i = (start_offset + start + min_len - 1) >>
5446                 PAGE_SHIFT;
5447
5448         if (start + min_len > eb->len) {
5449                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5450                        eb->start, eb->len, start, min_len);
5451                 return -EINVAL;
5452         }
5453
5454         if (i != end_i)
5455                 return 1;
5456
5457         if (i == 0) {
5458                 offset = start_offset;
5459                 *map_start = 0;
5460         } else {
5461                 offset = 0;
5462                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5463         }
5464
5465         p = eb->pages[i];
5466         kaddr = page_address(p);
5467         *map = kaddr + offset;
5468         *map_len = PAGE_SIZE - offset;
5469         return 0;
5470 }
5471
5472 int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
5473                          unsigned long start, unsigned long len)
5474 {
5475         size_t cur;
5476         size_t offset;
5477         struct page *page;
5478         char *kaddr;
5479         char *ptr = (char *)ptrv;
5480         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5481         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5482         int ret = 0;
5483
5484         WARN_ON(start > eb->len);
5485         WARN_ON(start + len > eb->start + eb->len);
5486
5487         offset = (start_offset + start) & (PAGE_SIZE - 1);
5488
5489         while (len > 0) {
5490                 page = eb->pages[i];
5491
5492                 cur = min(len, (PAGE_SIZE - offset));
5493
5494                 kaddr = page_address(page);
5495                 ret = memcmp(ptr, kaddr + offset, cur);
5496                 if (ret)
5497                         break;
5498
5499                 ptr += cur;
5500                 len -= cur;
5501                 offset = 0;
5502                 i++;
5503         }
5504         return ret;
5505 }
5506
5507 void write_extent_buffer_chunk_tree_uuid(struct extent_buffer *eb,
5508                 const void *srcv)
5509 {
5510         char *kaddr;
5511
5512         WARN_ON(!PageUptodate(eb->pages[0]));
5513         kaddr = page_address(eb->pages[0]);
5514         memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv,
5515                         BTRFS_FSID_SIZE);
5516 }
5517
5518 void write_extent_buffer_fsid(struct extent_buffer *eb, const void *srcv)
5519 {
5520         char *kaddr;
5521
5522         WARN_ON(!PageUptodate(eb->pages[0]));
5523         kaddr = page_address(eb->pages[0]);
5524         memcpy(kaddr + offsetof(struct btrfs_header, fsid), srcv,
5525                         BTRFS_FSID_SIZE);
5526 }
5527
5528 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5529                          unsigned long start, unsigned long len)
5530 {
5531         size_t cur;
5532         size_t offset;
5533         struct page *page;
5534         char *kaddr;
5535         char *src = (char *)srcv;
5536         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5537         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5538
5539         WARN_ON(start > eb->len);
5540         WARN_ON(start + len > eb->start + eb->len);
5541
5542         offset = (start_offset + start) & (PAGE_SIZE - 1);
5543
5544         while (len > 0) {
5545                 page = eb->pages[i];
5546                 WARN_ON(!PageUptodate(page));
5547
5548                 cur = min(len, PAGE_SIZE - offset);
5549                 kaddr = page_address(page);
5550                 memcpy(kaddr + offset, src, cur);
5551
5552                 src += cur;
5553                 len -= cur;
5554                 offset = 0;
5555                 i++;
5556         }
5557 }
5558
5559 void memzero_extent_buffer(struct extent_buffer *eb, unsigned long start,
5560                 unsigned long len)
5561 {
5562         size_t cur;
5563         size_t offset;
5564         struct page *page;
5565         char *kaddr;
5566         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5567         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5568
5569         WARN_ON(start > eb->len);
5570         WARN_ON(start + len > eb->start + eb->len);
5571
5572         offset = (start_offset + start) & (PAGE_SIZE - 1);
5573
5574         while (len > 0) {
5575                 page = eb->pages[i];
5576                 WARN_ON(!PageUptodate(page));
5577
5578                 cur = min(len, PAGE_SIZE - offset);
5579                 kaddr = page_address(page);
5580                 memset(kaddr + offset, 0, cur);
5581
5582                 len -= cur;
5583                 offset = 0;
5584                 i++;
5585         }
5586 }
5587
5588 void copy_extent_buffer_full(struct extent_buffer *dst,
5589                              struct extent_buffer *src)
5590 {
5591         int i;
5592         unsigned num_pages;
5593
5594         ASSERT(dst->len == src->len);
5595
5596         num_pages = num_extent_pages(dst->start, dst->len);
5597         for (i = 0; i < num_pages; i++)
5598                 copy_page(page_address(dst->pages[i]),
5599                                 page_address(src->pages[i]));
5600 }
5601
5602 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5603                         unsigned long dst_offset, unsigned long src_offset,
5604                         unsigned long len)
5605 {
5606         u64 dst_len = dst->len;
5607         size_t cur;
5608         size_t offset;
5609         struct page *page;
5610         char *kaddr;
5611         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5612         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5613
5614         WARN_ON(src->len != dst_len);
5615
5616         offset = (start_offset + dst_offset) &
5617                 (PAGE_SIZE - 1);
5618
5619         while (len > 0) {
5620                 page = dst->pages[i];
5621                 WARN_ON(!PageUptodate(page));
5622
5623                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5624
5625                 kaddr = page_address(page);
5626                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5627
5628                 src_offset += cur;
5629                 len -= cur;
5630                 offset = 0;
5631                 i++;
5632         }
5633 }
5634
5635 void le_bitmap_set(u8 *map, unsigned int start, int len)
5636 {
5637         u8 *p = map + BIT_BYTE(start);
5638         const unsigned int size = start + len;
5639         int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5640         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);
5641
5642         while (len - bits_to_set >= 0) {
5643                 *p |= mask_to_set;
5644                 len -= bits_to_set;
5645                 bits_to_set = BITS_PER_BYTE;
5646                 mask_to_set = ~0;
5647                 p++;
5648         }
5649         if (len) {
5650                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5651                 *p |= mask_to_set;
5652         }
5653 }
5654
5655 void le_bitmap_clear(u8 *map, unsigned int start, int len)
5656 {
5657         u8 *p = map + BIT_BYTE(start);
5658         const unsigned int size = start + len;
5659         int bits_to_clear = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5660         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(start);
5661
5662         while (len - bits_to_clear >= 0) {
5663                 *p &= ~mask_to_clear;
5664                 len -= bits_to_clear;
5665                 bits_to_clear = BITS_PER_BYTE;
5666                 mask_to_clear = ~0;
5667                 p++;
5668         }
5669         if (len) {
5670                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5671                 *p &= ~mask_to_clear;
5672         }
5673 }
5674
5675 /*
5676  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5677  * given bit number
5678  * @eb: the extent buffer
5679  * @start: offset of the bitmap item in the extent buffer
5680  * @nr: bit number
5681  * @page_index: return index of the page in the extent buffer that contains the
5682  * given bit number
5683  * @page_offset: return offset into the page given by page_index
5684  *
5685  * This helper hides the ugliness of finding the byte in an extent buffer which
5686  * contains a given bit.
5687  */
5688 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5689                                     unsigned long start, unsigned long nr,
5690                                     unsigned long *page_index,
5691                                     size_t *page_offset)
5692 {
5693         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5694         size_t byte_offset = BIT_BYTE(nr);
5695         size_t offset;
5696
5697         /*
5698          * The byte we want is the offset of the extent buffer + the offset of
5699          * the bitmap item in the extent buffer + the offset of the byte in the
5700          * bitmap item.
5701          */
5702         offset = start_offset + start + byte_offset;
5703
5704         *page_index = offset >> PAGE_SHIFT;
5705         *page_offset = offset & (PAGE_SIZE - 1);
5706 }
5707
5708 /**
5709  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5710  * @eb: the extent buffer
5711  * @start: offset of the bitmap item in the extent buffer
5712  * @nr: bit number to test
5713  */
5714 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5715                            unsigned long nr)
5716 {
5717         u8 *kaddr;
5718         struct page *page;
5719         unsigned long i;
5720         size_t offset;
5721
5722         eb_bitmap_offset(eb, start, nr, &i, &offset);
5723         page = eb->pages[i];
5724         WARN_ON(!PageUptodate(page));
5725         kaddr = page_address(page);
5726         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5727 }
5728
5729 /**
5730  * extent_buffer_bitmap_set - set an area of a bitmap
5731  * @eb: the extent buffer
5732  * @start: offset of the bitmap item in the extent buffer
5733  * @pos: bit number of the first bit
5734  * @len: number of bits to set
5735  */
5736 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5737                               unsigned long pos, unsigned long len)
5738 {
5739         u8 *kaddr;
5740         struct page *page;
5741         unsigned long i;
5742         size_t offset;
5743         const unsigned int size = pos + len;
5744         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5745         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5746
5747         eb_bitmap_offset(eb, start, pos, &i, &offset);
5748         page = eb->pages[i];
5749         WARN_ON(!PageUptodate(page));
5750         kaddr = page_address(page);
5751
5752         while (len >= bits_to_set) {
5753                 kaddr[offset] |= mask_to_set;
5754                 len -= bits_to_set;
5755                 bits_to_set = BITS_PER_BYTE;
5756                 mask_to_set = ~0;
5757                 if (++offset >= PAGE_SIZE && len > 0) {
5758                         offset = 0;
5759                         page = eb->pages[++i];
5760                         WARN_ON(!PageUptodate(page));
5761                         kaddr = page_address(page);
5762                 }
5763         }
5764         if (len) {
5765                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5766                 kaddr[offset] |= mask_to_set;
5767         }
5768 }
5769
5770
5771 /**
5772  * extent_buffer_bitmap_clear - clear an area of a bitmap
5773  * @eb: the extent buffer
5774  * @start: offset of the bitmap item in the extent buffer
5775  * @pos: bit number of the first bit
5776  * @len: number of bits to clear
5777  */
5778 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5779                                 unsigned long pos, unsigned long len)
5780 {
5781         u8 *kaddr;
5782         struct page *page;
5783         unsigned long i;
5784         size_t offset;
5785         const unsigned int size = pos + len;
5786         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5787         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5788
5789         eb_bitmap_offset(eb, start, pos, &i, &offset);
5790         page = eb->pages[i];
5791         WARN_ON(!PageUptodate(page));
5792         kaddr = page_address(page);
5793
5794         while (len >= bits_to_clear) {
5795                 kaddr[offset] &= ~mask_to_clear;
5796                 len -= bits_to_clear;
5797                 bits_to_clear = BITS_PER_BYTE;
5798                 mask_to_clear = ~0;
5799                 if (++offset >= PAGE_SIZE && len > 0) {
5800                         offset = 0;
5801                         page = eb->pages[++i];
5802                         WARN_ON(!PageUptodate(page));
5803                         kaddr = page_address(page);
5804                 }
5805         }
5806         if (len) {
5807                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5808                 kaddr[offset] &= ~mask_to_clear;
5809         }
5810 }
5811
5812 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5813 {
5814         unsigned long distance = (src > dst) ? src - dst : dst - src;
5815         return distance < len;
5816 }
5817
5818 static void copy_pages(struct page *dst_page, struct page *src_page,
5819                        unsigned long dst_off, unsigned long src_off,
5820                        unsigned long len)
5821 {
5822         char *dst_kaddr = page_address(dst_page);
5823         char *src_kaddr;
5824         int must_memmove = 0;
5825
5826         if (dst_page != src_page) {
5827                 src_kaddr = page_address(src_page);
5828         } else {
5829                 src_kaddr = dst_kaddr;
5830                 if (areas_overlap(src_off, dst_off, len))
5831                         must_memmove = 1;
5832         }
5833
5834         if (must_memmove)
5835                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5836         else
5837                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5838 }
5839
5840 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5841                            unsigned long src_offset, unsigned long len)
5842 {
5843         struct btrfs_fs_info *fs_info = dst->fs_info;
5844         size_t cur;
5845         size_t dst_off_in_page;
5846         size_t src_off_in_page;
5847         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5848         unsigned long dst_i;
5849         unsigned long src_i;
5850
5851         if (src_offset + len > dst->len) {
5852                 btrfs_err(fs_info,
5853                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5854                          src_offset, len, dst->len);
5855                 BUG_ON(1);
5856         }
5857         if (dst_offset + len > dst->len) {
5858                 btrfs_err(fs_info,
5859                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5860                          dst_offset, len, dst->len);
5861                 BUG_ON(1);
5862         }
5863
5864         while (len > 0) {
5865                 dst_off_in_page = (start_offset + dst_offset) &
5866                         (PAGE_SIZE - 1);
5867                 src_off_in_page = (start_offset + src_offset) &
5868                         (PAGE_SIZE - 1);
5869
5870                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5871                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5872
5873                 cur = min(len, (unsigned long)(PAGE_SIZE -
5874                                                src_off_in_page));
5875                 cur = min_t(unsigned long, cur,
5876                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5877
5878                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5879                            dst_off_in_page, src_off_in_page, cur);
5880
5881                 src_offset += cur;
5882                 dst_offset += cur;
5883                 len -= cur;
5884         }
5885 }
5886
5887 void memmove_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         unsigned long dst_end = dst_offset + len - 1;
5895         unsigned long src_end = src_offset + len - 1;
5896         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5897         unsigned long dst_i;
5898         unsigned long src_i;
5899
5900         if (src_offset + len > dst->len) {
5901                 btrfs_err(fs_info,
5902                           "memmove bogus src_offset %lu move len %lu len %lu",
5903                           src_offset, len, dst->len);
5904                 BUG_ON(1);
5905         }
5906         if (dst_offset + len > dst->len) {
5907                 btrfs_err(fs_info,
5908                           "memmove bogus dst_offset %lu move len %lu len %lu",
5909                           dst_offset, len, dst->len);
5910                 BUG_ON(1);
5911         }
5912         if (dst_offset < src_offset) {
5913                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5914                 return;
5915         }
5916         while (len > 0) {
5917                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5918                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5919
5920                 dst_off_in_page = (start_offset + dst_end) &
5921                         (PAGE_SIZE - 1);
5922                 src_off_in_page = (start_offset + src_end) &
5923                         (PAGE_SIZE - 1);
5924
5925                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5926                 cur = min(cur, dst_off_in_page + 1);
5927                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5928                            dst_off_in_page - cur + 1,
5929                            src_off_in_page - cur + 1, cur);
5930
5931                 dst_end -= cur;
5932                 src_end -= cur;
5933                 len -= cur;
5934         }
5935 }
5936
5937 int try_release_extent_buffer(struct page *page)
5938 {
5939         struct extent_buffer *eb;
5940
5941         /*
5942          * We need to make sure nobody is attaching this page to an eb right
5943          * now.
5944          */
5945         spin_lock(&page->mapping->private_lock);
5946         if (!PagePrivate(page)) {
5947                 spin_unlock(&page->mapping->private_lock);
5948                 return 1;
5949         }
5950
5951         eb = (struct extent_buffer *)page->private;
5952         BUG_ON(!eb);
5953
5954         /*
5955          * This is a little awful but should be ok, we need to make sure that
5956          * the eb doesn't disappear out from under us while we're looking at
5957          * this page.
5958          */
5959         spin_lock(&eb->refs_lock);
5960         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5961                 spin_unlock(&eb->refs_lock);
5962                 spin_unlock(&page->mapping->private_lock);
5963                 return 0;
5964         }
5965         spin_unlock(&page->mapping->private_lock);
5966
5967         /*
5968          * If tree ref isn't set then we know the ref on this eb is a real ref,
5969          * so just return, this page will likely be freed soon anyway.
5970          */
5971         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5972                 spin_unlock(&eb->refs_lock);
5973                 return 0;
5974         }
5975
5976         return release_extent_buffer(eb);
5977 }