fs/btrfs/extent_io.c

   1 #include <linux/bitops.h>
   2 #include <linux/slab.h>
   3 #include <linux/bio.h>
   4 #include <linux/mm.h>
   5 #include <linux/pagemap.h>
   6 #include <linux/page-flags.h>
   7 #include <linux/spinlock.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/swap.h>
  10 #include <linux/writeback.h>
  11 #include <linux/pagevec.h>
  12 #include <linux/prefetch.h>
  13 #include <linux/cleancache.h>
  14 #include "extent_io.h"
  15 #include "extent_map.h"
  16 #include "ctree.h"
  17 #include "btrfs_inode.h"
  18 #include "volumes.h"
  19 #include "check-integrity.h"
  20 #include "locking.h"
  21 #include "rcu-string.h"
  22 #include "backref.h"
  23
  24 static struct kmem_cache *extent_state_cache;
  25 static struct kmem_cache *extent_buffer_cache;
  26 static struct bio_set *btrfs_bioset;
  27
  28 #ifdef CONFIG_BTRFS_DEBUG
  29 static LIST_HEAD(buffers);
  30 static LIST_HEAD(states);
  31
  32 static DEFINE_SPINLOCK(leak_lock);
  33
  34 static inline
  35 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  36 {
  37         unsigned long flags;
  38
  39         spin_lock_irqsave(&leak_lock, flags);
  40         list_add(new, head);
  41         spin_unlock_irqrestore(&leak_lock, flags);
  42 }
  43
  44 static inline
  45 void btrfs_leak_debug_del(struct list_head *entry)
  46 {
  47         unsigned long flags;
  48
  49         spin_lock_irqsave(&leak_lock, flags);
  50         list_del(entry);
  51         spin_unlock_irqrestore(&leak_lock, flags);
  52 }
  53
  54 static inline
  55 void btrfs_leak_debug_check(void)
  56 {
  57         struct extent_state *state;
  58         struct extent_buffer *eb;
  59
  60         while (!list_empty(&states)) {
  61                 state = list_entry(states.next, struct extent_state, leak_list);
  62                 printk(KERN_ERR "BTRFS: state leak: start %llu end %llu "
  63                        "state %lu in tree %p refs %d\n",
  64                        state->start, state->end, state->state, state->tree,
  65                        atomic_read(&state->refs));
  66                 list_del(&state->leak_list);
  67                 kmem_cache_free(extent_state_cache, state);
  68         }
  69
  70         while (!list_empty(&buffers)) {
  71                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  72                 printk(KERN_ERR "BTRFS: buffer leak start %llu len %lu "
  73                        "refs %d\n",
  74                        eb->start, eb->len, atomic_read(&eb->refs));
  75                 list_del(&eb->leak_list);
  76                 kmem_cache_free(extent_buffer_cache, eb);
  77         }
  78 }
  79
  80 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  81         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  82 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  83                 struct extent_io_tree *tree, u64 start, u64 end)
  84 {
  85         struct inode *inode;
  86         u64 isize;
  87
  88         if (!tree->mapping)
  89                 return;
  90
  91         inode = tree->mapping->host;
  92         isize = i_size_read(inode);
  93         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
  94                 printk_ratelimited(KERN_DEBUG
  95                     "BTRFS: %s: ino %llu isize %llu odd range [%llu,%llu]\n",
  96                                 caller, btrfs_ino(inode), isize, start, end);
  97         }
  98 }
  99 #else
 100 #define btrfs_leak_debug_add(new, head) do {} while (0)
 101 #define btrfs_leak_debug_del(entry)     do {} while (0)
 102 #define btrfs_leak_debug_check()        do {} while (0)
 103 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 104 #endif
 105
 106 #define BUFFER_LRU_MAX 64
 107
 108 struct tree_entry {
 109         u64 start;
 110         u64 end;
 111         struct rb_node rb_node;
 112 };
 113
 114 struct extent_page_data {
 115         struct bio *bio;
 116         struct extent_io_tree *tree;
 117         get_extent_t *get_extent;
 118         unsigned long bio_flags;
 119
 120         /* tells writepage not to lock the state bits for this range
 121          * it still does the unlocking
 122          */
 123         unsigned int extent_locked:1;
 124
 125         /* tells the submit_bio code to use a WRITE_SYNC */
 126         unsigned int sync_io:1;
 127 };
 128
 129 static noinline void flush_write_bio(void *data);
 130 static inline struct btrfs_fs_info *
 131 tree_fs_info(struct extent_io_tree *tree)
 132 {
 133         if (!tree->mapping)
 134                 return NULL;
 135         return btrfs_sb(tree->mapping->host->i_sb);
 136 }
 137
 138 int __init extent_io_init(void)
 139 {
 140         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 141                         sizeof(struct extent_state), 0,
 142                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 143         if (!extent_state_cache)
 144                 return -ENOMEM;
 145
 146         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 147                         sizeof(struct extent_buffer), 0,
 148                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 149         if (!extent_buffer_cache)
 150                 goto free_state_cache;
 151
 152         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 153                                      offsetof(struct btrfs_io_bio, bio));
 154         if (!btrfs_bioset)
 155                 goto free_buffer_cache;
 156
 157         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 158                 goto free_bioset;
 159
 160         return 0;
 161
 162 free_bioset:
 163         bioset_free(btrfs_bioset);
 164         btrfs_bioset = NULL;
 165
 166 free_buffer_cache:
 167         kmem_cache_destroy(extent_buffer_cache);
 168         extent_buffer_cache = NULL;
 169
 170 free_state_cache:
 171         kmem_cache_destroy(extent_state_cache);
 172         extent_state_cache = NULL;
 173         return -ENOMEM;
 174 }
 175
 176 void extent_io_exit(void)
 177 {
 178         btrfs_leak_debug_check();
 179
 180         /*
 181          * Make sure all delayed rcu free are flushed before we
 182          * destroy caches.
 183          */
 184         rcu_barrier();
 185         if (extent_state_cache)
 186                 kmem_cache_destroy(extent_state_cache);
 187         if (extent_buffer_cache)
 188                 kmem_cache_destroy(extent_buffer_cache);
 189         if (btrfs_bioset)
 190                 bioset_free(btrfs_bioset);
 191 }
 192
 193 void extent_io_tree_init(struct extent_io_tree *tree,
 194                          struct address_space *mapping)
 195 {
 196         tree->state = RB_ROOT;
 197         tree->ops = NULL;
 198         tree->dirty_bytes = 0;
 199         spin_lock_init(&tree->lock);
 200         tree->mapping = mapping;
 201 }
 202
 203 static struct extent_state *alloc_extent_state(gfp_t mask)
 204 {
 205         struct extent_state *state;
 206
 207         state = kmem_cache_alloc(extent_state_cache, mask);
 208         if (!state)
 209                 return state;
 210         state->state = 0;
 211         state->private = 0;
 212         state->tree = NULL;
 213         btrfs_leak_debug_add(&state->leak_list, &states);
 214         atomic_set(&state->refs, 1);
 215         init_waitqueue_head(&state->wq);
 216         trace_alloc_extent_state(state, mask, _RET_IP_);
 217         return state;
 218 }
 219
 220 void free_extent_state(struct extent_state *state)
 221 {
 222         if (!state)
 223                 return;
 224         if (atomic_dec_and_test(&state->refs)) {
 225                 WARN_ON(state->tree);
 226                 btrfs_leak_debug_del(&state->leak_list);
 227                 trace_free_extent_state(state, _RET_IP_);
 228                 kmem_cache_free(extent_state_cache, state);
 229         }
 230 }
 231
 232 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
 233                                    struct rb_node *node,
 234                                    struct rb_node ***p_in,
 235                                    struct rb_node **parent_in)
 236 {
 237         struct rb_node **p = &root->rb_node;
 238         struct rb_node *parent = NULL;
 239         struct tree_entry *entry;
 240
 241         if (p_in && parent_in) {
 242                 p = *p_in;
 243                 parent = *parent_in;
 244                 goto do_insert;
 245         }
 246
 247         while (*p) {
 248                 parent = *p;
 249                 entry = rb_entry(parent, struct tree_entry, rb_node);
 250
 251                 if (offset < entry->start)
 252                         p = &(*p)->rb_left;
 253                 else if (offset > entry->end)
 254                         p = &(*p)->rb_right;
 255                 else
 256                         return parent;
 257         }
 258
 259 do_insert:
 260         rb_link_node(node, parent, p);
 261         rb_insert_color(node, root);
 262         return NULL;
 263 }
 264
 265 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 266                                       struct rb_node **prev_ret,
 267                                       struct rb_node **next_ret,
 268                                       struct rb_node ***p_ret,
 269                                       struct rb_node **parent_ret)
 270 {
 271         struct rb_root *root = &tree->state;
 272         struct rb_node **n = &root->rb_node;
 273         struct rb_node *prev = NULL;
 274         struct rb_node *orig_prev = NULL;
 275         struct tree_entry *entry;
 276         struct tree_entry *prev_entry = NULL;
 277
 278         while (*n) {
 279                 prev = *n;
 280                 entry = rb_entry(prev, struct tree_entry, rb_node);
 281                 prev_entry = entry;
 282
 283                 if (offset < entry->start)
 284                         n = &(*n)->rb_left;
 285                 else if (offset > entry->end)
 286                         n = &(*n)->rb_right;
 287                 else
 288                         return *n;
 289         }
 290
 291         if (p_ret)
 292                 *p_ret = n;
 293         if (parent_ret)
 294                 *parent_ret = prev;
 295
 296         if (prev_ret) {
 297                 orig_prev = prev;
 298                 while (prev && offset > prev_entry->end) {
 299                         prev = rb_next(prev);
 300                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 301                 }
 302                 *prev_ret = prev;
 303                 prev = orig_prev;
 304         }
 305
 306         if (next_ret) {
 307                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 308                 while (prev && offset < prev_entry->start) {
 309                         prev = rb_prev(prev);
 310                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 311                 }
 312                 *next_ret = prev;
 313         }
 314         return NULL;
 315 }
 316
 317 static inline struct rb_node *
 318 tree_search_for_insert(struct extent_io_tree *tree,
 319                        u64 offset,
 320                        struct rb_node ***p_ret,
 321                        struct rb_node **parent_ret)
 322 {
 323         struct rb_node *prev = NULL;
 324         struct rb_node *ret;
 325
 326         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 327         if (!ret)
 328                 return prev;
 329         return ret;
 330 }
 331
 332 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 333                                           u64 offset)
 334 {
 335         return tree_search_for_insert(tree, offset, NULL, NULL);
 336 }
 337
 338 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 339                      struct extent_state *other)
 340 {
 341         if (tree->ops && tree->ops->merge_extent_hook)
 342                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 343                                              other);
 344 }
 345
 346 /*
 347  * utility function to look for merge candidates inside a given range.
 348  * Any extents with matching state are merged together into a single
 349  * extent in the tree.  Extents with EXTENT_IO in their state field
 350  * are not merged because the end_io handlers need to be able to do
 351  * operations on them without sleeping (or doing allocations/splits).
 352  *
 353  * This should be called with the tree lock held.
 354  */
 355 static void merge_state(struct extent_io_tree *tree,
 356                         struct extent_state *state)
 357 {
 358         struct extent_state *other;
 359         struct rb_node *other_node;
 360
 361         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 362                 return;
 363
 364         other_node = rb_prev(&state->rb_node);
 365         if (other_node) {
 366                 other = rb_entry(other_node, struct extent_state, rb_node);
 367                 if (other->end == state->start - 1 &&
 368                     other->state == state->state) {
 369                         merge_cb(tree, state, other);
 370                         state->start = other->start;
 371                         other->tree = NULL;
 372                         rb_erase(&other->rb_node, &tree->state);
 373                         free_extent_state(other);
 374                 }
 375         }
 376         other_node = rb_next(&state->rb_node);
 377         if (other_node) {
 378                 other = rb_entry(other_node, struct extent_state, rb_node);
 379                 if (other->start == state->end + 1 &&
 380                     other->state == state->state) {
 381                         merge_cb(tree, state, other);
 382                         state->end = other->end;
 383                         other->tree = NULL;
 384                         rb_erase(&other->rb_node, &tree->state);
 385                         free_extent_state(other);
 386                 }
 387         }
 388 }
 389
 390 static void set_state_cb(struct extent_io_tree *tree,
 391                          struct extent_state *state, unsigned long *bits)
 392 {
 393         if (tree->ops && tree->ops->set_bit_hook)
 394                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 395 }
 396
 397 static void clear_state_cb(struct extent_io_tree *tree,
 398                            struct extent_state *state, unsigned long *bits)
 399 {
 400         if (tree->ops && tree->ops->clear_bit_hook)
 401                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 402 }
 403
 404 static void set_state_bits(struct extent_io_tree *tree,
 405                            struct extent_state *state, unsigned long *bits);
 406
 407 /*
 408  * insert an extent_state struct into the tree.  'bits' are set on the
 409  * struct before it is inserted.
 410  *
 411  * This may return -EEXIST if the extent is already there, in which case the
 412  * state struct is freed.
 413  *
 414  * The tree lock is not taken internally.  This is a utility function and
 415  * probably isn't what you want to call (see set/clear_extent_bit).
 416  */
 417 static int insert_state(struct extent_io_tree *tree,
 418                         struct extent_state *state, u64 start, u64 end,
 419                         struct rb_node ***p,
 420                         struct rb_node **parent,
 421                         unsigned long *bits)
 422 {
 423         struct rb_node *node;
 424
 425         if (end < start)
 426                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 427                        end, start);
 428         state->start = start;
 429         state->end = end;
 430
 431         set_state_bits(tree, state, bits);
 432
 433         node = tree_insert(&tree->state, end, &state->rb_node, p, parent);
 434         if (node) {
 435                 struct extent_state *found;
 436                 found = rb_entry(node, struct extent_state, rb_node);
 437                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
 438                        "%llu %llu\n",
 439                        found->start, found->end, start, end);
 440                 return -EEXIST;
 441         }
 442         state->tree = tree;
 443         merge_state(tree, state);
 444         return 0;
 445 }
 446
 447 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 448                      u64 split)
 449 {
 450         if (tree->ops && tree->ops->split_extent_hook)
 451                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 452 }
 453
 454 /*
 455  * split a given extent state struct in two, inserting the preallocated
 456  * struct 'prealloc' as the newly created second half.  'split' indicates an
 457  * offset inside 'orig' where it should be split.
 458  *
 459  * Before calling,
 460  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 461  * are two extent state structs in the tree:
 462  * prealloc: [orig->start, split - 1]
 463  * orig: [ split, orig->end ]
 464  *
 465  * The tree locks are not taken by this function. They need to be held
 466  * by the caller.
 467  */
 468 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 469                        struct extent_state *prealloc, u64 split)
 470 {
 471         struct rb_node *node;
 472
 473         split_cb(tree, orig, split);
 474
 475         prealloc->start = orig->start;
 476         prealloc->end = split - 1;
 477         prealloc->state = orig->state;
 478         orig->start = split;
 479
 480         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node,
 481                            NULL, NULL);
 482         if (node) {
 483                 free_extent_state(prealloc);
 484                 return -EEXIST;
 485         }
 486         prealloc->tree = tree;
 487         return 0;
 488 }
 489
 490 static struct extent_state *next_state(struct extent_state *state)
 491 {
 492         struct rb_node *next = rb_next(&state->rb_node);
 493         if (next)
 494                 return rb_entry(next, struct extent_state, rb_node);
 495         else
 496                 return NULL;
 497 }
 498
 499 /*
 500  * utility function to clear some bits in an extent state struct.
 501  * it will optionally wake up any one waiting on this state (wake == 1).
 502  *
 503  * If no bits are set on the state struct after clearing things, the
 504  * struct is freed and removed from the tree
 505  */
 506 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 507                                             struct extent_state *state,
 508                                             unsigned long *bits, int wake)
 509 {
 510         struct extent_state *next;
 511         unsigned long bits_to_clear = *bits & ~EXTENT_CTLBITS;
 512
 513         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 514                 u64 range = state->end - state->start + 1;
 515                 WARN_ON(range > tree->dirty_bytes);
 516                 tree->dirty_bytes -= range;
 517         }
 518         clear_state_cb(tree, state, bits);
 519         state->state &= ~bits_to_clear;
 520         if (wake)
 521                 wake_up(&state->wq);
 522         if (state->state == 0) {
 523                 next = next_state(state);
 524                 if (state->tree) {
 525                         rb_erase(&state->rb_node, &tree->state);
 526                         state->tree = NULL;
 527                         free_extent_state(state);
 528                 } else {
 529                         WARN_ON(1);
 530                 }
 531         } else {
 532                 merge_state(tree, state);
 533                 next = next_state(state);
 534         }
 535         return next;
 536 }
 537
 538 static struct extent_state *
 539 alloc_extent_state_atomic(struct extent_state *prealloc)
 540 {
 541         if (!prealloc)
 542                 prealloc = alloc_extent_state(GFP_ATOMIC);
 543
 544         return prealloc;
 545 }
 546
 547 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 548 {
 549         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 550                     "Extent tree was modified by another "
 551                     "thread while locked.");
 552 }
 553
 554 /*
 555  * clear some bits on a range in the tree.  This may require splitting
 556  * or inserting elements in the tree, so the gfp mask is used to
 557  * indicate which allocations or sleeping are allowed.
 558  *
 559  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 560  * the given range from the tree regardless of state (ie for truncate).
 561  *
 562  * the range [start, end] is inclusive.
 563  *
 564  * This takes the tree lock, and returns 0 on success and < 0 on error.
 565  */
 566 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 567                      unsigned long bits, int wake, int delete,
 568                      struct extent_state **cached_state,
 569                      gfp_t mask)
 570 {
 571         struct extent_state *state;
 572         struct extent_state *cached;
 573         struct extent_state *prealloc = NULL;
 574         struct rb_node *node;
 575         u64 last_end;
 576         int err;
 577         int clear = 0;
 578
 579         btrfs_debug_check_extent_io_range(tree, start, end);
 580
 581         if (bits & EXTENT_DELALLOC)
 582                 bits |= EXTENT_NORESERVE;
 583
 584         if (delete)
 585                 bits |= ~EXTENT_CTLBITS;
 586         bits |= EXTENT_FIRST_DELALLOC;
 587
 588         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 589                 clear = 1;
 590 again:
 591         if (!prealloc && (mask & __GFP_WAIT)) {
 592                 prealloc = alloc_extent_state(mask);
 593                 if (!prealloc)
 594                         return -ENOMEM;
 595         }
 596
 597         spin_lock(&tree->lock);
 598         if (cached_state) {
 599                 cached = *cached_state;
 600
 601                 if (clear) {
 602                         *cached_state = NULL;
 603                         cached_state = NULL;
 604                 }
 605
 606                 if (cached && cached->tree && cached->start <= start &&
 607                     cached->end > start) {
 608                         if (clear)
 609                                 atomic_dec(&cached->refs);
 610                         state = cached;
 611                         goto hit_next;
 612                 }
 613                 if (clear)
 614                         free_extent_state(cached);
 615         }
 616         /*
 617          * this search will find the extents that end after
 618          * our range starts
 619          */
 620         node = tree_search(tree, start);
 621         if (!node)
 622                 goto out;
 623         state = rb_entry(node, struct extent_state, rb_node);
 624 hit_next:
 625         if (state->start > end)
 626                 goto out;
 627         WARN_ON(state->end < start);
 628         last_end = state->end;
 629
 630         /* the state doesn't have the wanted bits, go ahead */
 631         if (!(state->state & bits)) {
 632                 state = next_state(state);
 633                 goto next;
 634         }
 635
 636         /*
 637          *     | ---- desired range ---- |
 638          *  | state | or
 639          *  | ------------- state -------------- |
 640          *
 641          * We need to split the extent we found, and may flip
 642          * bits on second half.
 643          *
 644          * If the extent we found extends past our range, we
 645          * just split and search again.  It'll get split again
 646          * the next time though.
 647          *
 648          * If the extent we found is inside our range, we clear
 649          * the desired bit on it.
 650          */
 651
 652         if (state->start < start) {
 653                 prealloc = alloc_extent_state_atomic(prealloc);
 654                 BUG_ON(!prealloc);
 655                 err = split_state(tree, state, prealloc, start);
 656                 if (err)
 657                         extent_io_tree_panic(tree, err);
 658
 659                 prealloc = NULL;
 660                 if (err)
 661                         goto out;
 662                 if (state->end <= end) {
 663                         state = clear_state_bit(tree, state, &bits, wake);
 664                         goto next;
 665                 }
 666                 goto search_again;
 667         }
 668         /*
 669          * | ---- desired range ---- |
 670          *                        | state |
 671          * We need to split the extent, and clear the bit
 672          * on the first half
 673          */
 674         if (state->start <= end && state->end > end) {
 675                 prealloc = alloc_extent_state_atomic(prealloc);
 676                 BUG_ON(!prealloc);
 677                 err = split_state(tree, state, prealloc, end + 1);
 678                 if (err)
 679                         extent_io_tree_panic(tree, err);
 680
 681                 if (wake)
 682                         wake_up(&state->wq);
 683
 684                 clear_state_bit(tree, prealloc, &bits, wake);
 685
 686                 prealloc = NULL;
 687                 goto out;
 688         }
 689
 690         state = clear_state_bit(tree, state, &bits, wake);
 691 next:
 692         if (last_end == (u64)-1)
 693                 goto out;
 694         start = last_end + 1;
 695         if (start <= end && state && !need_resched())
 696                 goto hit_next;
 697         goto search_again;
 698
 699 out:
 700         spin_unlock(&tree->lock);
 701         if (prealloc)
 702                 free_extent_state(prealloc);
 703
 704         return 0;
 705
 706 search_again:
 707         if (start > end)
 708                 goto out;
 709         spin_unlock(&tree->lock);
 710         if (mask & __GFP_WAIT)
 711                 cond_resched();
 712         goto again;
 713 }
 714
 715 static void wait_on_state(struct extent_io_tree *tree,
 716                           struct extent_state *state)
 717                 __releases(tree->lock)
 718                 __acquires(tree->lock)
 719 {
 720         DEFINE_WAIT(wait);
 721         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 722         spin_unlock(&tree->lock);
 723         schedule();
 724         spin_lock(&tree->lock);
 725         finish_wait(&state->wq, &wait);
 726 }
 727
 728 /*
 729  * waits for one or more bits to clear on a range in the state tree.
 730  * The range [start, end] is inclusive.
 731  * The tree lock is taken by this function
 732  */
 733 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 734                             unsigned long bits)
 735 {
 736         struct extent_state *state;
 737         struct rb_node *node;
 738
 739         btrfs_debug_check_extent_io_range(tree, start, end);
 740
 741         spin_lock(&tree->lock);
 742 again:
 743         while (1) {
 744                 /*
 745                  * this search will find all the extents that end after
 746                  * our range starts
 747                  */
 748                 node = tree_search(tree, start);
 749                 if (!node)
 750                         break;
 751
 752                 state = rb_entry(node, struct extent_state, rb_node);
 753
 754                 if (state->start > end)
 755                         goto out;
 756
 757                 if (state->state & bits) {
 758                         start = state->start;
 759                         atomic_inc(&state->refs);
 760                         wait_on_state(tree, state);
 761                         free_extent_state(state);
 762                         goto again;
 763                 }
 764                 start = state->end + 1;
 765
 766                 if (start > end)
 767                         break;
 768
 769                 cond_resched_lock(&tree->lock);
 770         }
 771 out:
 772         spin_unlock(&tree->lock);
 773 }
 774
 775 static void set_state_bits(struct extent_io_tree *tree,
 776                            struct extent_state *state,
 777                            unsigned long *bits)
 778 {
 779         unsigned long bits_to_set = *bits & ~EXTENT_CTLBITS;
 780
 781         set_state_cb(tree, state, bits);
 782         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 783                 u64 range = state->end - state->start + 1;
 784                 tree->dirty_bytes += range;
 785         }
 786         state->state |= bits_to_set;
 787 }
 788
 789 static void cache_state(struct extent_state *state,
 790                         struct extent_state **cached_ptr)
 791 {
 792         if (cached_ptr && !(*cached_ptr)) {
 793                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 794                         *cached_ptr = state;
 795                         atomic_inc(&state->refs);
 796                 }
 797         }
 798 }
 799
 800 /*
 801  * set some bits on a range in the tree.  This may require allocations or
 802  * sleeping, so the gfp mask is used to indicate what is allowed.
 803  *
 804  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 805  * part of the range already has the desired bits set.  The start of the
 806  * existing range is returned in failed_start in this case.
 807  *
 808  * [start, end] is inclusive This takes the tree lock.
 809  */
 810
 811 static int __must_check
 812 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 813                  unsigned long bits, unsigned long exclusive_bits,
 814                  u64 *failed_start, struct extent_state **cached_state,
 815                  gfp_t mask)
 816 {
 817         struct extent_state *state;
 818         struct extent_state *prealloc = NULL;
 819         struct rb_node *node;
 820         struct rb_node **p;
 821         struct rb_node *parent;
 822         int err = 0;
 823         u64 last_start;
 824         u64 last_end;
 825
 826         btrfs_debug_check_extent_io_range(tree, start, end);
 827
 828         bits |= EXTENT_FIRST_DELALLOC;
 829 again:
 830         if (!prealloc && (mask & __GFP_WAIT)) {
 831                 prealloc = alloc_extent_state(mask);
 832                 BUG_ON(!prealloc);
 833         }
 834
 835         spin_lock(&tree->lock);
 836         if (cached_state && *cached_state) {
 837                 state = *cached_state;
 838                 if (state->start <= start && state->end > start &&
 839                     state->tree) {
 840                         node = &state->rb_node;
 841                         goto hit_next;
 842                 }
 843         }
 844         /*
 845          * this search will find all the extents that end after
 846          * our range starts.
 847          */
 848         node = tree_search_for_insert(tree, start, &p, &parent);
 849         if (!node) {
 850                 prealloc = alloc_extent_state_atomic(prealloc);
 851                 BUG_ON(!prealloc);
 852                 err = insert_state(tree, prealloc, start, end,
 853                                    &p, &parent, &bits);
 854                 if (err)
 855                         extent_io_tree_panic(tree, err);
 856
 857                 cache_state(prealloc, cached_state);
 858                 prealloc = NULL;
 859                 goto out;
 860         }
 861         state = rb_entry(node, struct extent_state, rb_node);
 862 hit_next:
 863         last_start = state->start;
 864         last_end = state->end;
 865
 866         /*
 867          * | ---- desired range ---- |
 868          * | state |
 869          *
 870          * Just lock what we found and keep going
 871          */
 872         if (state->start == start && state->end <= end) {
 873                 if (state->state & exclusive_bits) {
 874                         *failed_start = state->start;
 875                         err = -EEXIST;
 876                         goto out;
 877                 }
 878
 879                 set_state_bits(tree, state, &bits);
 880                 cache_state(state, cached_state);
 881                 merge_state(tree, state);
 882                 if (last_end == (u64)-1)
 883                         goto out;
 884                 start = last_end + 1;
 885                 state = next_state(state);
 886                 if (start < end && state && state->start == start &&
 887                     !need_resched())
 888                         goto hit_next;
 889                 goto search_again;
 890         }
 891
 892         /*
 893          *     | ---- desired range ---- |
 894          * | state |
 895          *   or
 896          * | ------------- state -------------- |
 897          *
 898          * We need to split the extent we found, and may flip bits on
 899          * second half.
 900          *
 901          * If the extent we found extends past our
 902          * range, we just split and search again.  It'll get split
 903          * again the next time though.
 904          *
 905          * If the extent we found is inside our range, we set the
 906          * desired bit on it.
 907          */
 908         if (state->start < start) {
 909                 if (state->state & exclusive_bits) {
 910                         *failed_start = start;
 911                         err = -EEXIST;
 912                         goto out;
 913                 }
 914
 915                 prealloc = alloc_extent_state_atomic(prealloc);
 916                 BUG_ON(!prealloc);
 917                 err = split_state(tree, state, prealloc, start);
 918                 if (err)
 919                         extent_io_tree_panic(tree, err);
 920
 921                 prealloc = NULL;
 922                 if (err)
 923                         goto out;
 924                 if (state->end <= end) {
 925                         set_state_bits(tree, state, &bits);
 926                         cache_state(state, cached_state);
 927                         merge_state(tree, state);
 928                         if (last_end == (u64)-1)
 929                                 goto out;
 930                         start = last_end + 1;
 931                         state = next_state(state);
 932                         if (start < end && state && state->start == start &&
 933                             !need_resched())
 934                                 goto hit_next;
 935                 }
 936                 goto search_again;
 937         }
 938         /*
 939          * | ---- desired range ---- |
 940          *     | state | or               | state |
 941          *
 942          * There's a hole, we need to insert something in it and
 943          * ignore the extent we found.
 944          */
 945         if (state->start > start) {
 946                 u64 this_end;
 947                 if (end < last_start)
 948                         this_end = end;
 949                 else
 950                         this_end = last_start - 1;
 951
 952                 prealloc = alloc_extent_state_atomic(prealloc);
 953                 BUG_ON(!prealloc);
 954
 955                 /*
 956                  * Avoid to free 'prealloc' if it can be merged with
 957                  * the later extent.
 958                  */
 959                 err = insert_state(tree, prealloc, start, this_end,
 960                                    NULL, NULL, &bits);
 961                 if (err)
 962                         extent_io_tree_panic(tree, err);
 963
 964                 cache_state(prealloc, cached_state);
 965                 prealloc = NULL;
 966                 start = this_end + 1;
 967                 goto search_again;
 968         }
 969         /*
 970          * | ---- desired range ---- |
 971          *                        | state |
 972          * We need to split the extent, and set the bit
 973          * on the first half
 974          */
 975         if (state->start <= end && state->end > end) {
 976                 if (state->state & exclusive_bits) {
 977                         *failed_start = start;
 978                         err = -EEXIST;
 979                         goto out;
 980                 }
 981
 982                 prealloc = alloc_extent_state_atomic(prealloc);
 983                 BUG_ON(!prealloc);
 984                 err = split_state(tree, state, prealloc, end + 1);
 985                 if (err)
 986                         extent_io_tree_panic(tree, err);
 987
 988                 set_state_bits(tree, prealloc, &bits);
 989                 cache_state(prealloc, cached_state);
 990                 merge_state(tree, prealloc);
 991                 prealloc = NULL;
 992                 goto out;
 993         }
 994
 995         goto search_again;
 996
 997 out:
 998         spin_unlock(&tree->lock);
 999         if (prealloc)
1000                 free_extent_state(prealloc);
1001
1002         return err;
1003
1004 search_again:
1005         if (start > end)
1006                 goto out;
1007         spin_unlock(&tree->lock);
1008         if (mask & __GFP_WAIT)
1009                 cond_resched();
1010         goto again;
1011 }
1012
1013 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1014                    unsigned long bits, u64 * failed_start,
1015                    struct extent_state **cached_state, gfp_t mask)
1016 {
1017         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1018                                 cached_state, mask);
1019 }
1020
1021
1022 /**
1023  * convert_extent_bit - convert all bits in a given range from one bit to
1024  *                      another
1025  * @tree:       the io tree to search
1026  * @start:      the start offset in bytes
1027  * @end:        the end offset in bytes (inclusive)
1028  * @bits:       the bits to set in this range
1029  * @clear_bits: the bits to clear in this range
1030  * @cached_state:       state that we're going to cache
1031  * @mask:       the allocation mask
1032  *
1033  * This will go through and set bits for the given range.  If any states exist
1034  * already in this range they are set with the given bit and cleared of the
1035  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1036  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1037  * boundary bits like LOCK.
1038  */
1039 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1040                        unsigned long bits, unsigned long clear_bits,
1041                        struct extent_state **cached_state, gfp_t mask)
1042 {
1043         struct extent_state *state;
1044         struct extent_state *prealloc = NULL;
1045         struct rb_node *node;
1046         struct rb_node **p;
1047         struct rb_node *parent;
1048         int err = 0;
1049         u64 last_start;
1050         u64 last_end;
1051
1052         btrfs_debug_check_extent_io_range(tree, start, end);
1053
1054 again:
1055         if (!prealloc && (mask & __GFP_WAIT)) {
1056                 prealloc = alloc_extent_state(mask);
1057                 if (!prealloc)
1058                         return -ENOMEM;
1059         }
1060
1061         spin_lock(&tree->lock);
1062         if (cached_state && *cached_state) {
1063                 state = *cached_state;
1064                 if (state->start <= start && state->end > start &&
1065                     state->tree) {
1066                         node = &state->rb_node;
1067                         goto hit_next;
1068                 }
1069         }
1070
1071         /*
1072          * this search will find all the extents that end after
1073          * our range starts.
1074          */
1075         node = tree_search_for_insert(tree, start, &p, &parent);
1076         if (!node) {
1077                 prealloc = alloc_extent_state_atomic(prealloc);
1078                 if (!prealloc) {
1079                         err = -ENOMEM;
1080                         goto out;
1081                 }
1082                 err = insert_state(tree, prealloc, start, end,
1083                                    &p, &parent, &bits);
1084                 if (err)
1085                         extent_io_tree_panic(tree, err);
1086                 cache_state(prealloc, cached_state);
1087                 prealloc = NULL;
1088                 goto out;
1089         }
1090         state = rb_entry(node, struct extent_state, rb_node);
1091 hit_next:
1092         last_start = state->start;
1093         last_end = state->end;
1094
1095         /*
1096          * | ---- desired range ---- |
1097          * | state |
1098          *
1099          * Just lock what we found and keep going
1100          */
1101         if (state->start == start && state->end <= end) {
1102                 set_state_bits(tree, state, &bits);
1103                 cache_state(state, cached_state);
1104                 state = clear_state_bit(tree, state, &clear_bits, 0);
1105                 if (last_end == (u64)-1)
1106                         goto out;
1107                 start = last_end + 1;
1108                 if (start < end && state && state->start == start &&
1109                     !need_resched())
1110                         goto hit_next;
1111                 goto search_again;
1112         }
1113
1114         /*
1115          *     | ---- desired range ---- |
1116          * | state |
1117          *   or
1118          * | ------------- state -------------- |
1119          *
1120          * We need to split the extent we found, and may flip bits on
1121          * second half.
1122          *
1123          * If the extent we found extends past our
1124          * range, we just split and search again.  It'll get split
1125          * again the next time though.
1126          *
1127          * If the extent we found is inside our range, we set the
1128          * desired bit on it.
1129          */
1130         if (state->start < start) {
1131                 prealloc = alloc_extent_state_atomic(prealloc);
1132                 if (!prealloc) {
1133                         err = -ENOMEM;
1134                         goto out;
1135                 }
1136                 err = split_state(tree, state, prealloc, start);
1137                 if (err)
1138                         extent_io_tree_panic(tree, err);
1139                 prealloc = NULL;
1140                 if (err)
1141                         goto out;
1142                 if (state->end <= end) {
1143                         set_state_bits(tree, state, &bits);
1144                         cache_state(state, cached_state);
1145                         state = clear_state_bit(tree, state, &clear_bits, 0);
1146                         if (last_end == (u64)-1)
1147                                 goto out;
1148                         start = last_end + 1;
1149                         if (start < end && state && state->start == start &&
1150                             !need_resched())
1151                                 goto hit_next;
1152                 }
1153                 goto search_again;
1154         }
1155         /*
1156          * | ---- desired range ---- |
1157          *     | state | or               | state |
1158          *
1159          * There's a hole, we need to insert something in it and
1160          * ignore the extent we found.
1161          */
1162         if (state->start > start) {
1163                 u64 this_end;
1164                 if (end < last_start)
1165                         this_end = end;
1166                 else
1167                         this_end = last_start - 1;
1168
1169                 prealloc = alloc_extent_state_atomic(prealloc);
1170                 if (!prealloc) {
1171                         err = -ENOMEM;
1172                         goto out;
1173                 }
1174
1175                 /*
1176                  * Avoid to free 'prealloc' if it can be merged with
1177                  * the later extent.
1178                  */
1179                 err = insert_state(tree, prealloc, start, this_end,
1180                                    NULL, NULL, &bits);
1181                 if (err)
1182                         extent_io_tree_panic(tree, err);
1183                 cache_state(prealloc, cached_state);
1184                 prealloc = NULL;
1185                 start = this_end + 1;
1186                 goto search_again;
1187         }
1188         /*
1189          * | ---- desired range ---- |
1190          *                        | state |
1191          * We need to split the extent, and set the bit
1192          * on the first half
1193          */
1194         if (state->start <= end && state->end > end) {
1195                 prealloc = alloc_extent_state_atomic(prealloc);
1196                 if (!prealloc) {
1197                         err = -ENOMEM;
1198                         goto out;
1199                 }
1200
1201                 err = split_state(tree, state, prealloc, end + 1);
1202                 if (err)
1203                         extent_io_tree_panic(tree, err);
1204
1205                 set_state_bits(tree, prealloc, &bits);
1206                 cache_state(prealloc, cached_state);
1207                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1208                 prealloc = NULL;
1209                 goto out;
1210         }
1211
1212         goto search_again;
1213
1214 out:
1215         spin_unlock(&tree->lock);
1216         if (prealloc)
1217                 free_extent_state(prealloc);
1218
1219         return err;
1220
1221 search_again:
1222         if (start > end)
1223                 goto out;
1224         spin_unlock(&tree->lock);
1225         if (mask & __GFP_WAIT)
1226                 cond_resched();
1227         goto again;
1228 }
1229
1230 /* wrappers around set/clear extent bit */
1231 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1232                      gfp_t mask)
1233 {
1234         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1235                               NULL, mask);
1236 }
1237
1238 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1239                     unsigned long bits, gfp_t mask)
1240 {
1241         return set_extent_bit(tree, start, end, bits, NULL,
1242                               NULL, mask);
1243 }
1244
1245 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1246                       unsigned long bits, gfp_t mask)
1247 {
1248         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1249 }
1250
1251 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1252                         struct extent_state **cached_state, gfp_t mask)
1253 {
1254         return set_extent_bit(tree, start, end,
1255                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1256                               NULL, cached_state, mask);
1257 }
1258
1259 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1260                       struct extent_state **cached_state, gfp_t mask)
1261 {
1262         return set_extent_bit(tree, start, end,
1263                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1264                               NULL, cached_state, mask);
1265 }
1266
1267 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1268                        gfp_t mask)
1269 {
1270         return clear_extent_bit(tree, start, end,
1271                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1272                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1273 }
1274
1275 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1276                      gfp_t mask)
1277 {
1278         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1279                               NULL, mask);
1280 }
1281
1282 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1283                         struct extent_state **cached_state, gfp_t mask)
1284 {
1285         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1286                               cached_state, mask);
1287 }
1288
1289 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1290                           struct extent_state **cached_state, gfp_t mask)
1291 {
1292         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1293                                 cached_state, mask);
1294 }
1295
1296 /*
1297  * either insert or lock state struct between start and end use mask to tell
1298  * us if waiting is desired.
1299  */
1300 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1301                      unsigned long bits, struct extent_state **cached_state)
1302 {
1303         int err;
1304         u64 failed_start;
1305         while (1) {
1306                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1307                                        EXTENT_LOCKED, &failed_start,
1308                                        cached_state, GFP_NOFS);
1309                 if (err == -EEXIST) {
1310                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1311                         start = failed_start;
1312                 } else
1313                         break;
1314                 WARN_ON(start > end);
1315         }
1316         return err;
1317 }
1318
1319 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1320 {
1321         return lock_extent_bits(tree, start, end, 0, NULL);
1322 }
1323
1324 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1325 {
1326         int err;
1327         u64 failed_start;
1328
1329         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1330                                &failed_start, NULL, GFP_NOFS);
1331         if (err == -EEXIST) {
1332                 if (failed_start > start)
1333                         clear_extent_bit(tree, start, failed_start - 1,
1334                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1335                 return 0;
1336         }
1337         return 1;
1338 }
1339
1340 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1341                          struct extent_state **cached, gfp_t mask)
1342 {
1343         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1344                                 mask);
1345 }
1346
1347 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1348 {
1349         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1350                                 GFP_NOFS);
1351 }
1352
1353 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1354 {
1355         unsigned long index = start >> PAGE_CACHE_SHIFT;
1356         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1357         struct page *page;
1358
1359         while (index <= end_index) {
1360                 page = find_get_page(inode->i_mapping, index);
1361                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1362                 clear_page_dirty_for_io(page);
1363                 page_cache_release(page);
1364                 index++;
1365         }
1366         return 0;
1367 }
1368
1369 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1370 {
1371         unsigned long index = start >> PAGE_CACHE_SHIFT;
1372         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1373         struct page *page;
1374
1375         while (index <= end_index) {
1376                 page = find_get_page(inode->i_mapping, index);
1377                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1378                 account_page_redirty(page);
1379                 __set_page_dirty_nobuffers(page);
1380                 page_cache_release(page);
1381                 index++;
1382         }
1383         return 0;
1384 }
1385
1386 /*
1387  * helper function to set both pages and extents in the tree writeback
1388  */
1389 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1390 {
1391         unsigned long index = start >> PAGE_CACHE_SHIFT;
1392         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1393         struct page *page;
1394
1395         while (index <= end_index) {
1396                 page = find_get_page(tree->mapping, index);
1397                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1398                 set_page_writeback(page);
1399                 page_cache_release(page);
1400                 index++;
1401         }
1402         return 0;
1403 }
1404
1405 /* find the first state struct with 'bits' set after 'start', and
1406  * return it.  tree->lock must be held.  NULL will returned if
1407  * nothing was found after 'start'
1408  */
1409 static struct extent_state *
1410 find_first_extent_bit_state(struct extent_io_tree *tree,
1411                             u64 start, unsigned long bits)
1412 {
1413         struct rb_node *node;
1414         struct extent_state *state;
1415
1416         /*
1417          * this search will find all the extents that end after
1418          * our range starts.
1419          */
1420         node = tree_search(tree, start);
1421         if (!node)
1422                 goto out;
1423
1424         while (1) {
1425                 state = rb_entry(node, struct extent_state, rb_node);
1426                 if (state->end >= start && (state->state & bits))
1427                         return state;
1428
1429                 node = rb_next(node);
1430                 if (!node)
1431                         break;
1432         }
1433 out:
1434         return NULL;
1435 }
1436
1437 /*
1438  * find the first offset in the io tree with 'bits' set. zero is
1439  * returned if we find something, and *start_ret and *end_ret are
1440  * set to reflect the state struct that was found.
1441  *
1442  * If nothing was found, 1 is returned. If found something, return 0.
1443  */
1444 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1445                           u64 *start_ret, u64 *end_ret, unsigned long bits,
1446                           struct extent_state **cached_state)
1447 {
1448         struct extent_state *state;
1449         struct rb_node *n;
1450         int ret = 1;
1451
1452         spin_lock(&tree->lock);
1453         if (cached_state && *cached_state) {
1454                 state = *cached_state;
1455                 if (state->end == start - 1 && state->tree) {
1456                         n = rb_next(&state->rb_node);
1457                         while (n) {
1458                                 state = rb_entry(n, struct extent_state,
1459                                                  rb_node);
1460                                 if (state->state & bits)
1461                                         goto got_it;
1462                                 n = rb_next(n);
1463                         }
1464                         free_extent_state(*cached_state);
1465                         *cached_state = NULL;
1466                         goto out;
1467                 }
1468                 free_extent_state(*cached_state);
1469                 *cached_state = NULL;
1470         }
1471
1472         state = find_first_extent_bit_state(tree, start, bits);
1473 got_it:
1474         if (state) {
1475                 cache_state(state, cached_state);
1476                 *start_ret = state->start;
1477                 *end_ret = state->end;
1478                 ret = 0;
1479         }
1480 out:
1481         spin_unlock(&tree->lock);
1482         return ret;
1483 }
1484
1485 /*
1486  * find a contiguous range of bytes in the file marked as delalloc, not
1487  * more than 'max_bytes'.  start and end are used to return the range,
1488  *
1489  * 1 is returned if we find something, 0 if nothing was in the tree
1490  */
1491 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1492                                         u64 *start, u64 *end, u64 max_bytes,
1493                                         struct extent_state **cached_state)
1494 {
1495         struct rb_node *node;
1496         struct extent_state *state;
1497         u64 cur_start = *start;
1498         u64 found = 0;
1499         u64 total_bytes = 0;
1500
1501         spin_lock(&tree->lock);
1502
1503         /*
1504          * this search will find all the extents that end after
1505          * our range starts.
1506          */
1507         node = tree_search(tree, cur_start);
1508         if (!node) {
1509                 if (!found)
1510                         *end = (u64)-1;
1511                 goto out;
1512         }
1513
1514         while (1) {
1515                 state = rb_entry(node, struct extent_state, rb_node);
1516                 if (found && (state->start != cur_start ||
1517                               (state->state & EXTENT_BOUNDARY))) {
1518                         goto out;
1519                 }
1520                 if (!(state->state & EXTENT_DELALLOC)) {
1521                         if (!found)
1522                                 *end = state->end;
1523                         goto out;
1524                 }
1525                 if (!found) {
1526                         *start = state->start;
1527                         *cached_state = state;
1528                         atomic_inc(&state->refs);
1529                 }
1530                 found++;
1531                 *end = state->end;
1532                 cur_start = state->end + 1;
1533                 node = rb_next(node);
1534                 total_bytes += state->end - state->start + 1;
1535                 if (total_bytes >= max_bytes)
1536                         break;
1537                 if (!node)
1538                         break;
1539         }
1540 out:
1541         spin_unlock(&tree->lock);
1542         return found;
1543 }
1544
1545 static noinline void __unlock_for_delalloc(struct inode *inode,
1546                                            struct page *locked_page,
1547                                            u64 start, u64 end)
1548 {
1549         int ret;
1550         struct page *pages[16];
1551         unsigned long index = start >> PAGE_CACHE_SHIFT;
1552         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1553         unsigned long nr_pages = end_index - index + 1;
1554         int i;
1555
1556         if (index == locked_page->index && end_index == index)
1557                 return;
1558
1559         while (nr_pages > 0) {
1560                 ret = find_get_pages_contig(inode->i_mapping, index,
1561                                      min_t(unsigned long, nr_pages,
1562                                      ARRAY_SIZE(pages)), pages);
1563                 for (i = 0; i < ret; i++) {
1564                         if (pages[i] != locked_page)
1565                                 unlock_page(pages[i]);
1566                         page_cache_release(pages[i]);
1567                 }
1568                 nr_pages -= ret;
1569                 index += ret;
1570                 cond_resched();
1571         }
1572 }
1573
1574 static noinline int lock_delalloc_pages(struct inode *inode,
1575                                         struct page *locked_page,
1576                                         u64 delalloc_start,
1577                                         u64 delalloc_end)
1578 {
1579         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1580         unsigned long start_index = index;
1581         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1582         unsigned long pages_locked = 0;
1583         struct page *pages[16];
1584         unsigned long nrpages;
1585         int ret;
1586         int i;
1587
1588         /* the caller is responsible for locking the start index */
1589         if (index == locked_page->index && index == end_index)
1590                 return 0;
1591
1592         /* skip the page at the start index */
1593         nrpages = end_index - index + 1;
1594         while (nrpages > 0) {
1595                 ret = find_get_pages_contig(inode->i_mapping, index,
1596                                      min_t(unsigned long,
1597                                      nrpages, ARRAY_SIZE(pages)), pages);
1598                 if (ret == 0) {
1599                         ret = -EAGAIN;
1600                         goto done;
1601                 }
1602                 /* now we have an array of pages, lock them all */
1603                 for (i = 0; i < ret; i++) {
1604                         /*
1605                          * the caller is taking responsibility for
1606                          * locked_page
1607                          */
1608                         if (pages[i] != locked_page) {
1609                                 lock_page(pages[i]);
1610                                 if (!PageDirty(pages[i]) ||
1611                                     pages[i]->mapping != inode->i_mapping) {
1612                                         ret = -EAGAIN;
1613                                         unlock_page(pages[i]);
1614                                         page_cache_release(pages[i]);
1615                                         goto done;
1616                                 }
1617                         }
1618                         page_cache_release(pages[i]);
1619                         pages_locked++;
1620                 }
1621                 nrpages -= ret;
1622                 index += ret;
1623                 cond_resched();
1624         }
1625         ret = 0;
1626 done:
1627         if (ret && pages_locked) {
1628                 __unlock_for_delalloc(inode, locked_page,
1629                               delalloc_start,
1630                               ((u64)(start_index + pages_locked - 1)) <<
1631                               PAGE_CACHE_SHIFT);
1632         }
1633         return ret;
1634 }
1635
1636 /*
1637  * find a contiguous range of bytes in the file marked as delalloc, not
1638  * more than 'max_bytes'.  start and end are used to return the range,
1639  *
1640  * 1 is returned if we find something, 0 if nothing was in the tree
1641  */
1642 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1643                                     struct extent_io_tree *tree,
1644                                     struct page *locked_page, u64 *start,
1645                                     u64 *end, u64 max_bytes)
1646 {
1647         u64 delalloc_start;
1648         u64 delalloc_end;
1649         u64 found;
1650         struct extent_state *cached_state = NULL;
1651         int ret;
1652         int loops = 0;
1653
1654 again:
1655         /* step one, find a bunch of delalloc bytes starting at start */
1656         delalloc_start = *start;
1657         delalloc_end = 0;
1658         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1659                                     max_bytes, &cached_state);
1660         if (!found || delalloc_end <= *start) {
1661                 *start = delalloc_start;
1662                 *end = delalloc_end;
1663                 free_extent_state(cached_state);
1664                 return 0;
1665         }
1666
1667         /*
1668          * start comes from the offset of locked_page.  We have to lock
1669          * pages in order, so we can't process delalloc bytes before
1670          * locked_page
1671          */
1672         if (delalloc_start < *start)
1673                 delalloc_start = *start;
1674
1675         /*
1676          * make sure to limit the number of pages we try to lock down
1677          */
1678         if (delalloc_end + 1 - delalloc_start > max_bytes)
1679                 delalloc_end = delalloc_start + max_bytes - 1;
1680
1681         /* step two, lock all the pages after the page that has start */
1682         ret = lock_delalloc_pages(inode, locked_page,
1683                                   delalloc_start, delalloc_end);
1684         if (ret == -EAGAIN) {
1685                 /* some of the pages are gone, lets avoid looping by
1686                  * shortening the size of the delalloc range we're searching
1687                  */
1688                 free_extent_state(cached_state);
1689                 cached_state = NULL;
1690                 if (!loops) {
1691                         max_bytes = PAGE_CACHE_SIZE;
1692                         loops = 1;
1693                         goto again;
1694                 } else {
1695                         found = 0;
1696                         goto out_failed;
1697                 }
1698         }
1699         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1700
1701         /* step three, lock the state bits for the whole range */
1702         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1703
1704         /* then test to make sure it is all still delalloc */
1705         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1706                              EXTENT_DELALLOC, 1, cached_state);
1707         if (!ret) {
1708                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1709                                      &cached_state, GFP_NOFS);
1710                 __unlock_for_delalloc(inode, locked_page,
1711                               delalloc_start, delalloc_end);
1712                 cond_resched();
1713                 goto again;
1714         }
1715         free_extent_state(cached_state);
1716         *start = delalloc_start;
1717         *end = delalloc_end;
1718 out_failed:
1719         return found;
1720 }
1721
1722 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1723                                  struct page *locked_page,
1724                                  unsigned long clear_bits,
1725                                  unsigned long page_ops)
1726 {
1727         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1728         int ret;
1729         struct page *pages[16];
1730         unsigned long index = start >> PAGE_CACHE_SHIFT;
1731         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1732         unsigned long nr_pages = end_index - index + 1;
1733         int i;
1734
1735         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1736         if (page_ops == 0)
1737                 return 0;
1738
1739         while (nr_pages > 0) {
1740                 ret = find_get_pages_contig(inode->i_mapping, index,
1741                                      min_t(unsigned long,
1742                                      nr_pages, ARRAY_SIZE(pages)), pages);
1743                 for (i = 0; i < ret; i++) {
1744
1745                         if (page_ops & PAGE_SET_PRIVATE2)
1746                                 SetPagePrivate2(pages[i]);
1747
1748                         if (pages[i] == locked_page) {
1749                                 page_cache_release(pages[i]);
1750                                 continue;
1751                         }
1752                         if (page_ops & PAGE_CLEAR_DIRTY)
1753                                 clear_page_dirty_for_io(pages[i]);
1754                         if (page_ops & PAGE_SET_WRITEBACK)
1755                                 set_page_writeback(pages[i]);
1756                         if (page_ops & PAGE_END_WRITEBACK)
1757                                 end_page_writeback(pages[i]);
1758                         if (page_ops & PAGE_UNLOCK)
1759                                 unlock_page(pages[i]);
1760                         page_cache_release(pages[i]);
1761                 }
1762                 nr_pages -= ret;
1763                 index += ret;
1764                 cond_resched();
1765         }
1766         return 0;
1767 }
1768
1769 /*
1770  * count the number of bytes in the tree that have a given bit(s)
1771  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1772  * cached.  The total number found is returned.
1773  */
1774 u64 count_range_bits(struct extent_io_tree *tree,
1775                      u64 *start, u64 search_end, u64 max_bytes,
1776                      unsigned long bits, int contig)
1777 {
1778         struct rb_node *node;
1779         struct extent_state *state;
1780         u64 cur_start = *start;
1781         u64 total_bytes = 0;
1782         u64 last = 0;
1783         int found = 0;
1784
1785         if (WARN_ON(search_end <= cur_start))
1786                 return 0;
1787
1788         spin_lock(&tree->lock);
1789         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1790                 total_bytes = tree->dirty_bytes;
1791                 goto out;
1792         }
1793         /*
1794          * this search will find all the extents that end after
1795          * our range starts.
1796          */
1797         node = tree_search(tree, cur_start);
1798         if (!node)
1799                 goto out;
1800
1801         while (1) {
1802                 state = rb_entry(node, struct extent_state, rb_node);
1803                 if (state->start > search_end)
1804                         break;
1805                 if (contig && found && state->start > last + 1)
1806                         break;
1807                 if (state->end >= cur_start && (state->state & bits) == bits) {
1808                         total_bytes += min(search_end, state->end) + 1 -
1809                                        max(cur_start, state->start);
1810                         if (total_bytes >= max_bytes)
1811                                 break;
1812                         if (!found) {
1813                                 *start = max(cur_start, state->start);
1814                                 found = 1;
1815                         }
1816                         last = state->end;
1817                 } else if (contig && found) {
1818                         break;
1819                 }
1820                 node = rb_next(node);
1821                 if (!node)
1822                         break;
1823         }
1824 out:
1825         spin_unlock(&tree->lock);
1826         return total_bytes;
1827 }
1828
1829 /*
1830  * set the private field for a given byte offset in the tree.  If there isn't
1831  * an extent_state there already, this does nothing.
1832  */
1833 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1834 {
1835         struct rb_node *node;
1836         struct extent_state *state;
1837         int ret = 0;
1838
1839         spin_lock(&tree->lock);
1840         /*
1841          * this search will find all the extents that end after
1842          * our range starts.
1843          */
1844         node = tree_search(tree, start);
1845         if (!node) {
1846                 ret = -ENOENT;
1847                 goto out;
1848         }
1849         state = rb_entry(node, struct extent_state, rb_node);
1850         if (state->start != start) {
1851                 ret = -ENOENT;
1852                 goto out;
1853         }
1854         state->private = private;
1855 out:
1856         spin_unlock(&tree->lock);
1857         return ret;
1858 }
1859
1860 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1861 {
1862         struct rb_node *node;
1863         struct extent_state *state;
1864         int ret = 0;
1865
1866         spin_lock(&tree->lock);
1867         /*
1868          * this search will find all the extents that end after
1869          * our range starts.
1870          */
1871         node = tree_search(tree, start);
1872         if (!node) {
1873                 ret = -ENOENT;
1874                 goto out;
1875         }
1876         state = rb_entry(node, struct extent_state, rb_node);
1877         if (state->start != start) {
1878                 ret = -ENOENT;
1879                 goto out;
1880         }
1881         *private = state->private;
1882 out:
1883         spin_unlock(&tree->lock);
1884         return ret;
1885 }
1886
1887 /*
1888  * searches a range in the state tree for a given mask.
1889  * If 'filled' == 1, this returns 1 only if every extent in the tree
1890  * has the bits set.  Otherwise, 1 is returned if any bit in the
1891  * range is found set.
1892  */
1893 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1894                    unsigned long bits, int filled, struct extent_state *cached)
1895 {
1896         struct extent_state *state = NULL;
1897         struct rb_node *node;
1898         int bitset = 0;
1899
1900         spin_lock(&tree->lock);
1901         if (cached && cached->tree && cached->start <= start &&
1902             cached->end > start)
1903                 node = &cached->rb_node;
1904         else
1905                 node = tree_search(tree, start);
1906         while (node && start <= end) {
1907                 state = rb_entry(node, struct extent_state, rb_node);
1908
1909                 if (filled && state->start > start) {
1910                         bitset = 0;
1911                         break;
1912                 }
1913
1914                 if (state->start > end)
1915                         break;
1916
1917                 if (state->state & bits) {
1918                         bitset = 1;
1919                         if (!filled)
1920                                 break;
1921                 } else if (filled) {
1922                         bitset = 0;
1923                         break;
1924                 }
1925
1926                 if (state->end == (u64)-1)
1927                         break;
1928
1929                 start = state->end + 1;
1930                 if (start > end)
1931                         break;
1932                 node = rb_next(node);
1933                 if (!node) {
1934                         if (filled)
1935                                 bitset = 0;
1936                         break;
1937                 }
1938         }
1939         spin_unlock(&tree->lock);
1940         return bitset;
1941 }
1942
1943 /*
1944  * helper function to set a given page up to date if all the
1945  * extents in the tree for that page are up to date
1946  */
1947 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1948 {
1949         u64 start = page_offset(page);
1950         u64 end = start + PAGE_CACHE_SIZE - 1;
1951         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1952                 SetPageUptodate(page);
1953 }
1954
1955 /*
1956  * When IO fails, either with EIO or csum verification fails, we
1957  * try other mirrors that might have a good copy of the data.  This
1958  * io_failure_record is used to record state as we go through all the
1959  * mirrors.  If another mirror has good data, the page is set up to date
1960  * and things continue.  If a good mirror can't be found, the original
1961  * bio end_io callback is called to indicate things have failed.
1962  */
1963 struct io_failure_record {
1964         struct page *page;
1965         u64 start;
1966         u64 len;
1967         u64 logical;
1968         unsigned long bio_flags;
1969         int this_mirror;
1970         int failed_mirror;
1971         int in_validation;
1972 };
1973
1974 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1975                                 int did_repair)
1976 {
1977         int ret;
1978         int err = 0;
1979         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1980
1981         set_state_private(failure_tree, rec->start, 0);
1982         ret = clear_extent_bits(failure_tree, rec->start,
1983                                 rec->start + rec->len - 1,
1984                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1985         if (ret)
1986                 err = ret;
1987
1988         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1989                                 rec->start + rec->len - 1,
1990                                 EXTENT_DAMAGED, GFP_NOFS);
1991         if (ret && !err)
1992                 err = ret;
1993
1994         kfree(rec);
1995         return err;
1996 }
1997
1998 /*
1999  * this bypasses the standard btrfs submit functions deliberately, as
2000  * the standard behavior is to write all copies in a raid setup. here we only
2001  * want to write the one bad copy. so we do the mapping for ourselves and issue
2002  * submit_bio directly.
2003  * to avoid any synchronization issues, wait for the data after writing, which
2004  * actually prevents the read that triggered the error from finishing.
2005  * currently, there can be no more than two copies of every data bit. thus,
2006  * exactly one rewrite is required.
2007  */
2008 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 start,
2009                         u64 length, u64 logical, struct page *page,
2010                         int mirror_num)
2011 {
2012         struct bio *bio;
2013         struct btrfs_device *dev;
2014         u64 map_length = 0;
2015         u64 sector;
2016         struct btrfs_bio *bbio = NULL;
2017         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2018         int ret;
2019
2020         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2021         BUG_ON(!mirror_num);
2022
2023         /* we can't repair anything in raid56 yet */
2024         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2025                 return 0;
2026
2027         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2028         if (!bio)
2029                 return -EIO;
2030         bio->bi_iter.bi_size = 0;
2031         map_length = length;
2032
2033         ret = btrfs_map_block(fs_info, WRITE, logical,
2034                               &map_length, &bbio, mirror_num);
2035         if (ret) {
2036                 bio_put(bio);
2037                 return -EIO;
2038         }
2039         BUG_ON(mirror_num != bbio->mirror_num);
2040         sector = bbio->stripes[mirror_num-1].physical >> 9;
2041         bio->bi_iter.bi_sector = sector;
2042         dev = bbio->stripes[mirror_num-1].dev;
2043         kfree(bbio);
2044         if (!dev || !dev->bdev || !dev->writeable) {
2045                 bio_put(bio);
2046                 return -EIO;
2047         }
2048         bio->bi_bdev = dev->bdev;
2049         bio_add_page(bio, page, length, start - page_offset(page));
2050
2051         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio)) {
2052                 /* try to remap that extent elsewhere? */
2053                 bio_put(bio);
2054                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2055                 return -EIO;
2056         }
2057
2058         printk_ratelimited_in_rcu(KERN_INFO
2059                         "BTRFS: read error corrected: ino %lu off %llu "
2060                     "(dev %s sector %llu)\n", page->mapping->host->i_ino,
2061                     start, rcu_str_deref(dev->name), sector);
2062
2063         bio_put(bio);
2064         return 0;
2065 }
2066
2067 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2068                          int mirror_num)
2069 {
2070         u64 start = eb->start;
2071         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2072         int ret = 0;
2073
2074         if (root->fs_info->sb->s_flags & MS_RDONLY)
2075                 return -EROFS;
2076
2077         for (i = 0; i < num_pages; i++) {
2078                 struct page *p = extent_buffer_page(eb, i);
2079                 ret = repair_io_failure(root->fs_info, start, PAGE_CACHE_SIZE,
2080                                         start, p, mirror_num);
2081                 if (ret)
2082                         break;
2083                 start += PAGE_CACHE_SIZE;
2084         }
2085
2086         return ret;
2087 }
2088
2089 /*
2090  * each time an IO finishes, we do a fast check in the IO failure tree
2091  * to see if we need to process or clean up an io_failure_record
2092  */
2093 static int clean_io_failure(u64 start, struct page *page)
2094 {
2095         u64 private;
2096         u64 private_failure;
2097         struct io_failure_record *failrec;
2098         struct inode *inode = page->mapping->host;
2099         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2100         struct extent_state *state;
2101         int num_copies;
2102         int did_repair = 0;
2103         int ret;
2104
2105         private = 0;
2106         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2107                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2108         if (!ret)
2109                 return 0;
2110
2111         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2112                                 &private_failure);
2113         if (ret)
2114                 return 0;
2115
2116         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2117         BUG_ON(!failrec->this_mirror);
2118
2119         if (failrec->in_validation) {
2120                 /* there was no real error, just free the record */
2121                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2122                          failrec->start);
2123                 did_repair = 1;
2124                 goto out;
2125         }
2126         if (fs_info->sb->s_flags & MS_RDONLY)
2127                 goto out;
2128
2129         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2130         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2131                                             failrec->start,
2132                                             EXTENT_LOCKED);
2133         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2134
2135         if (state && state->start <= failrec->start &&
2136             state->end >= failrec->start + failrec->len - 1) {
2137                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2138                                               failrec->len);
2139                 if (num_copies > 1)  {
2140                         ret = repair_io_failure(fs_info, start, failrec->len,
2141                                                 failrec->logical, page,
2142                                                 failrec->failed_mirror);
2143                         did_repair = !ret;
2144                 }
2145                 ret = 0;
2146         }
2147
2148 out:
2149         if (!ret)
2150                 ret = free_io_failure(inode, failrec, did_repair);
2151
2152         return ret;
2153 }
2154
2155 /*
2156  * this is a generic handler for readpage errors (default
2157  * readpage_io_failed_hook). if other copies exist, read those and write back
2158  * good data to the failed position. does not investigate in remapping the
2159  * failed extent elsewhere, hoping the device will be smart enough to do this as
2160  * needed
2161  */
2162
2163 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2164                               struct page *page, u64 start, u64 end,
2165                               int failed_mirror)
2166 {
2167         struct io_failure_record *failrec = NULL;
2168         u64 private;
2169         struct extent_map *em;
2170         struct inode *inode = page->mapping->host;
2171         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2172         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2173         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2174         struct bio *bio;
2175         struct btrfs_io_bio *btrfs_failed_bio;
2176         struct btrfs_io_bio *btrfs_bio;
2177         int num_copies;
2178         int ret;
2179         int read_mode;
2180         u64 logical;
2181
2182         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2183
2184         ret = get_state_private(failure_tree, start, &private);
2185         if (ret) {
2186                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2187                 if (!failrec)
2188                         return -ENOMEM;
2189                 failrec->start = start;
2190                 failrec->len = end - start + 1;
2191                 failrec->this_mirror = 0;
2192                 failrec->bio_flags = 0;
2193                 failrec->in_validation = 0;
2194
2195                 read_lock(&em_tree->lock);
2196                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2197                 if (!em) {
2198                         read_unlock(&em_tree->lock);
2199                         kfree(failrec);
2200                         return -EIO;
2201                 }
2202
2203                 if (em->start > start || em->start + em->len <= start) {
2204                         free_extent_map(em);
2205                         em = NULL;
2206                 }
2207                 read_unlock(&em_tree->lock);
2208
2209                 if (!em) {
2210                         kfree(failrec);
2211                         return -EIO;
2212                 }
2213                 logical = start - em->start;
2214                 logical = em->block_start + logical;
2215                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2216                         logical = em->block_start;
2217                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2218                         extent_set_compress_type(&failrec->bio_flags,
2219                                                  em->compress_type);
2220                 }
2221                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2222                          "len=%llu\n", logical, start, failrec->len);
2223                 failrec->logical = logical;
2224                 free_extent_map(em);
2225
2226                 /* set the bits in the private failure tree */
2227                 ret = set_extent_bits(failure_tree, start, end,
2228                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2229                 if (ret >= 0)
2230                         ret = set_state_private(failure_tree, start,
2231                                                 (u64)(unsigned long)failrec);
2232                 /* set the bits in the inode's tree */
2233                 if (ret >= 0)
2234                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2235                                                 GFP_NOFS);
2236                 if (ret < 0) {
2237                         kfree(failrec);
2238                         return ret;
2239                 }
2240         } else {
2241                 failrec = (struct io_failure_record *)(unsigned long)private;
2242                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2243                          "start=%llu, len=%llu, validation=%d\n",
2244                          failrec->logical, failrec->start, failrec->len,
2245                          failrec->in_validation);
2246                 /*
2247                  * when data can be on disk more than twice, add to failrec here
2248                  * (e.g. with a list for failed_mirror) to make
2249                  * clean_io_failure() clean all those errors at once.
2250                  */
2251         }
2252         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2253                                       failrec->logical, failrec->len);
2254         if (num_copies == 1) {
2255                 /*
2256                  * we only have a single copy of the data, so don't bother with
2257                  * all the retry and error correction code that follows. no
2258                  * matter what the error is, it is very likely to persist.
2259                  */
2260                 pr_debug("bio_readpage_error: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2261                          num_copies, failrec->this_mirror, failed_mirror);
2262                 free_io_failure(inode, failrec, 0);
2263                 return -EIO;
2264         }
2265
2266         /*
2267          * there are two premises:
2268          *      a) deliver good data to the caller
2269          *      b) correct the bad sectors on disk
2270          */
2271         if (failed_bio->bi_vcnt > 1) {
2272                 /*
2273                  * to fulfill b), we need to know the exact failing sectors, as
2274                  * we don't want to rewrite any more than the failed ones. thus,
2275                  * we need separate read requests for the failed bio
2276                  *
2277                  * if the following BUG_ON triggers, our validation request got
2278                  * merged. we need separate requests for our algorithm to work.
2279                  */
2280                 BUG_ON(failrec->in_validation);
2281                 failrec->in_validation = 1;
2282                 failrec->this_mirror = failed_mirror;
2283                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2284         } else {
2285                 /*
2286                  * we're ready to fulfill a) and b) alongside. get a good copy
2287                  * of the failed sector and if we succeed, we have setup
2288                  * everything for repair_io_failure to do the rest for us.
2289                  */
2290                 if (failrec->in_validation) {
2291                         BUG_ON(failrec->this_mirror != failed_mirror);
2292                         failrec->in_validation = 0;
2293                         failrec->this_mirror = 0;
2294                 }
2295                 failrec->failed_mirror = failed_mirror;
2296                 failrec->this_mirror++;
2297                 if (failrec->this_mirror == failed_mirror)
2298                         failrec->this_mirror++;
2299                 read_mode = READ_SYNC;
2300         }
2301
2302         if (failrec->this_mirror > num_copies) {
2303                 pr_debug("bio_readpage_error: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2304                          num_copies, failrec->this_mirror, failed_mirror);
2305                 free_io_failure(inode, failrec, 0);
2306                 return -EIO;
2307         }
2308
2309         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2310         if (!bio) {
2311                 free_io_failure(inode, failrec, 0);
2312                 return -EIO;
2313         }
2314         bio->bi_end_io = failed_bio->bi_end_io;
2315         bio->bi_iter.bi_sector = failrec->logical >> 9;
2316         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2317         bio->bi_iter.bi_size = 0;
2318
2319         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2320         if (btrfs_failed_bio->csum) {
2321                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2322                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2323
2324                 btrfs_bio = btrfs_io_bio(bio);
2325                 btrfs_bio->csum = btrfs_bio->csum_inline;
2326                 phy_offset >>= inode->i_sb->s_blocksize_bits;
2327                 phy_offset *= csum_size;
2328                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + phy_offset,
2329                        csum_size);
2330         }
2331
2332         bio_add_page(bio, page, failrec->len, start - page_offset(page));
2333
2334         pr_debug("bio_readpage_error: submitting new read[%#x] to "
2335                  "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2336                  failrec->this_mirror, num_copies, failrec->in_validation);
2337
2338         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2339                                          failrec->this_mirror,
2340                                          failrec->bio_flags, 0);
2341         return ret;
2342 }
2343
2344 /* lots and lots of room for performance fixes in the end_bio funcs */
2345
2346 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2347 {
2348         int uptodate = (err == 0);
2349         struct extent_io_tree *tree;
2350         int ret = 0;
2351
2352         tree = &BTRFS_I(page->mapping->host)->io_tree;
2353
2354         if (tree->ops && tree->ops->writepage_end_io_hook) {
2355                 ret = tree->ops->writepage_end_io_hook(page, start,
2356                                                end, NULL, uptodate);
2357                 if (ret)
2358                         uptodate = 0;
2359         }
2360
2361         if (!uptodate) {
2362                 ClearPageUptodate(page);
2363                 SetPageError(page);
2364                 ret = ret < 0 ? ret : -EIO;
2365                 mapping_set_error(page->mapping, ret);
2366         }
2367         return 0;
2368 }
2369
2370 /*
2371  * after a writepage IO is done, we need to:
2372  * clear the uptodate bits on error
2373  * clear the writeback bits in the extent tree for this IO
2374  * end_page_writeback if the page has no more pending IO
2375  *
2376  * Scheduling is not allowed, so the extent state tree is expected
2377  * to have one and only one object corresponding to this IO.
2378  */
2379 static void end_bio_extent_writepage(struct bio *bio, int err)
2380 {
2381         struct bio_vec *bvec;
2382         u64 start;
2383         u64 end;
2384         int i;
2385
2386         bio_for_each_segment_all(bvec, bio, i) {
2387                 struct page *page = bvec->bv_page;
2388
2389                 /* We always issue full-page reads, but if some block
2390                  * in a page fails to read, blk_update_request() will
2391                  * advance bv_offset and adjust bv_len to compensate.
2392                  * Print a warning for nonzero offsets, and an error
2393                  * if they don't add up to a full page.  */
2394                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2395                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2396                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2397                                    "partial page write in btrfs with offset %u and length %u",
2398                                         bvec->bv_offset, bvec->bv_len);
2399                         else
2400                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2401                                    "incomplete page write in btrfs with offset %u and "
2402                                    "length %u",
2403                                         bvec->bv_offset, bvec->bv_len);
2404                 }
2405
2406                 start = page_offset(page);
2407                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2408
2409                 if (end_extent_writepage(page, err, start, end))
2410                         continue;
2411
2412                 end_page_writeback(page);
2413         }
2414
2415         bio_put(bio);
2416 }
2417
2418 static void
2419 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2420                               int uptodate)
2421 {
2422         struct extent_state *cached = NULL;
2423         u64 end = start + len - 1;
2424
2425         if (uptodate && tree->track_uptodate)
2426                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2427         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2428 }
2429
2430 /*
2431  * after a readpage IO is done, we need to:
2432  * clear the uptodate bits on error
2433  * set the uptodate bits if things worked
2434  * set the page up to date if all extents in the tree are uptodate
2435  * clear the lock bit in the extent tree
2436  * unlock the page if there are no other extents locked for it
2437  *
2438  * Scheduling is not allowed, so the extent state tree is expected
2439  * to have one and only one object corresponding to this IO.
2440  */
2441 static void end_bio_extent_readpage(struct bio *bio, int err)
2442 {
2443         struct bio_vec *bvec;
2444         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2445         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2446         struct extent_io_tree *tree;
2447         u64 offset = 0;
2448         u64 start;
2449         u64 end;
2450         u64 len;
2451         u64 extent_start = 0;
2452         u64 extent_len = 0;
2453         int mirror;
2454         int ret;
2455         int i;
2456
2457         if (err)
2458                 uptodate = 0;
2459
2460         bio_for_each_segment_all(bvec, bio, i) {
2461                 struct page *page = bvec->bv_page;
2462                 struct inode *inode = page->mapping->host;
2463
2464                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2465                          "mirror=%lu\n", (u64)bio->bi_iter.bi_sector, err,
2466                          io_bio->mirror_num);
2467                 tree = &BTRFS_I(inode)->io_tree;
2468
2469                 /* We always issue full-page reads, but if some block
2470                  * in a page fails to read, blk_update_request() will
2471                  * advance bv_offset and adjust bv_len to compensate.
2472                  * Print a warning for nonzero offsets, and an error
2473                  * if they don't add up to a full page.  */
2474                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2475                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2476                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2477                                    "partial page read in btrfs with offset %u and length %u",
2478                                         bvec->bv_offset, bvec->bv_len);
2479                         else
2480                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2481                                    "incomplete page read in btrfs with offset %u and "
2482                                    "length %u",
2483                                         bvec->bv_offset, bvec->bv_len);
2484                 }
2485
2486                 start = page_offset(page);
2487                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2488                 len = bvec->bv_len;
2489
2490                 mirror = io_bio->mirror_num;
2491                 if (likely(uptodate && tree->ops &&
2492                            tree->ops->readpage_end_io_hook)) {
2493                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2494                                                               page, start, end,
2495                                                               mirror);
2496                         if (ret)
2497                                 uptodate = 0;
2498                         else
2499                                 clean_io_failure(start, page);
2500                 }
2501
2502                 if (likely(uptodate))
2503                         goto readpage_ok;
2504
2505                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2506                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2507                         if (!ret && !err &&
2508                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2509                                 uptodate = 1;
2510                 } else {
2511                         /*
2512                          * The generic bio_readpage_error handles errors the
2513                          * following way: If possible, new read requests are
2514                          * created and submitted and will end up in
2515                          * end_bio_extent_readpage as well (if we're lucky, not
2516                          * in the !uptodate case). In that case it returns 0 and
2517                          * we just go on with the next page in our bio. If it
2518                          * can't handle the error it will return -EIO and we
2519                          * remain responsible for that page.
2520                          */
2521                         ret = bio_readpage_error(bio, offset, page, start, end,
2522                                                  mirror);
2523                         if (ret == 0) {
2524                                 uptodate =
2525                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2526                                 if (err)
2527                                         uptodate = 0;
2528                                 offset += len;
2529                                 continue;
2530                         }
2531                 }
2532 readpage_ok:
2533                 if (likely(uptodate)) {
2534                         loff_t i_size = i_size_read(inode);
2535                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2536                         unsigned offset;
2537
2538                         /* Zero out the end if this page straddles i_size */
2539                         offset = i_size & (PAGE_CACHE_SIZE-1);
2540                         if (page->index == end_index && offset)
2541                                 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2542                         SetPageUptodate(page);
2543                 } else {
2544                         ClearPageUptodate(page);
2545                         SetPageError(page);
2546                 }
2547                 unlock_page(page);
2548                 offset += len;
2549
2550                 if (unlikely(!uptodate)) {
2551                         if (extent_len) {
2552                                 endio_readpage_release_extent(tree,
2553                                                               extent_start,
2554                                                               extent_len, 1);
2555                                 extent_start = 0;
2556                                 extent_len = 0;
2557                         }
2558                         endio_readpage_release_extent(tree, start,
2559                                                       end - start + 1, 0);
2560                 } else if (!extent_len) {
2561                         extent_start = start;
2562                         extent_len = end + 1 - start;
2563                 } else if (extent_start + extent_len == start) {
2564                         extent_len += end + 1 - start;
2565                 } else {
2566                         endio_readpage_release_extent(tree, extent_start,
2567                                                       extent_len, uptodate);
2568                         extent_start = start;
2569                         extent_len = end + 1 - start;
2570                 }
2571         }
2572
2573         if (extent_len)
2574                 endio_readpage_release_extent(tree, extent_start, extent_len,
2575                                               uptodate);
2576         if (io_bio->end_io)
2577                 io_bio->end_io(io_bio, err);
2578         bio_put(bio);
2579 }
2580
2581 /*
2582  * this allocates from the btrfs_bioset.  We're returning a bio right now
2583  * but you can call btrfs_io_bio for the appropriate container_of magic
2584  */
2585 struct bio *
2586 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2587                 gfp_t gfp_flags)
2588 {
2589         struct btrfs_io_bio *btrfs_bio;
2590         struct bio *bio;
2591
2592         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2593
2594         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2595                 while (!bio && (nr_vecs /= 2)) {
2596                         bio = bio_alloc_bioset(gfp_flags,
2597                                                nr_vecs, btrfs_bioset);
2598                 }
2599         }
2600
2601         if (bio) {
2602                 bio->bi_bdev = bdev;
2603                 bio->bi_iter.bi_sector = first_sector;
2604                 btrfs_bio = btrfs_io_bio(bio);
2605                 btrfs_bio->csum = NULL;
2606                 btrfs_bio->csum_allocated = NULL;
2607                 btrfs_bio->end_io = NULL;
2608         }
2609         return bio;
2610 }
2611
2612 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2613 {
2614         return bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2615 }
2616
2617
2618 /* this also allocates from the btrfs_bioset */
2619 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2620 {
2621         struct btrfs_io_bio *btrfs_bio;
2622         struct bio *bio;
2623
2624         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2625         if (bio) {
2626                 btrfs_bio = btrfs_io_bio(bio);
2627                 btrfs_bio->csum = NULL;
2628                 btrfs_bio->csum_allocated = NULL;
2629                 btrfs_bio->end_io = NULL;
2630         }
2631         return bio;
2632 }
2633
2634
2635 static int __must_check submit_one_bio(int rw, struct bio *bio,
2636                                        int mirror_num, unsigned long bio_flags)
2637 {
2638         int ret = 0;
2639         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2640         struct page *page = bvec->bv_page;
2641         struct extent_io_tree *tree = bio->bi_private;
2642         u64 start;
2643
2644         start = page_offset(page) + bvec->bv_offset;
2645
2646         bio->bi_private = NULL;
2647
2648         bio_get(bio);
2649
2650         if (tree->ops && tree->ops->submit_bio_hook)
2651                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2652                                            mirror_num, bio_flags, start);
2653         else
2654                 btrfsic_submit_bio(rw, bio);
2655
2656         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2657                 ret = -EOPNOTSUPP;
2658         bio_put(bio);
2659         return ret;
2660 }
2661
2662 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2663                      unsigned long offset, size_t size, struct bio *bio,
2664                      unsigned long bio_flags)
2665 {
2666         int ret = 0;
2667         if (tree->ops && tree->ops->merge_bio_hook)
2668                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2669                                                 bio_flags);
2670         BUG_ON(ret < 0);
2671         return ret;
2672
2673 }
2674
2675 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2676                               struct page *page, sector_t sector,
2677                               size_t size, unsigned long offset,
2678                               struct block_device *bdev,
2679                               struct bio **bio_ret,
2680                               unsigned long max_pages,
2681                               bio_end_io_t end_io_func,
2682                               int mirror_num,
2683                               unsigned long prev_bio_flags,
2684                               unsigned long bio_flags)
2685 {
2686         int ret = 0;
2687         struct bio *bio;
2688         int nr;
2689         int contig = 0;
2690         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2691         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2692         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2693
2694         if (bio_ret && *bio_ret) {
2695                 bio = *bio_ret;
2696                 if (old_compressed)
2697                         contig = bio->bi_iter.bi_sector == sector;
2698                 else
2699                         contig = bio_end_sector(bio) == sector;
2700
2701                 if (prev_bio_flags != bio_flags || !contig ||
2702                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2703                     bio_add_page(bio, page, page_size, offset) < page_size) {
2704                         ret = submit_one_bio(rw, bio, mirror_num,
2705                                              prev_bio_flags);
2706                         if (ret < 0)
2707                                 return ret;
2708                         bio = NULL;
2709                 } else {
2710                         return 0;
2711                 }
2712         }
2713         if (this_compressed)
2714                 nr = BIO_MAX_PAGES;
2715         else
2716                 nr = bio_get_nr_vecs(bdev);
2717
2718         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2719         if (!bio)
2720                 return -ENOMEM;
2721
2722         bio_add_page(bio, page, page_size, offset);
2723         bio->bi_end_io = end_io_func;
2724         bio->bi_private = tree;
2725
2726         if (bio_ret)
2727                 *bio_ret = bio;
2728         else
2729                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2730
2731         return ret;
2732 }
2733
2734 static void attach_extent_buffer_page(struct extent_buffer *eb,
2735                                       struct page *page)
2736 {
2737         if (!PagePrivate(page)) {
2738                 SetPagePrivate(page);
2739                 page_cache_get(page);
2740                 set_page_private(page, (unsigned long)eb);
2741         } else {
2742                 WARN_ON(page->private != (unsigned long)eb);
2743         }
2744 }
2745
2746 void set_page_extent_mapped(struct page *page)
2747 {
2748         if (!PagePrivate(page)) {
2749                 SetPagePrivate(page);
2750                 page_cache_get(page);
2751                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2752         }
2753 }
2754
2755 static struct extent_map *
2756 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2757                  u64 start, u64 len, get_extent_t *get_extent,
2758                  struct extent_map **em_cached)
2759 {
2760         struct extent_map *em;
2761
2762         if (em_cached && *em_cached) {
2763                 em = *em_cached;
2764                 if (em->in_tree && start >= em->start &&
2765                     start < extent_map_end(em)) {
2766                         atomic_inc(&em->refs);
2767                         return em;
2768                 }
2769
2770                 free_extent_map(em);
2771                 *em_cached = NULL;
2772         }
2773
2774         em = get_extent(inode, page, pg_offset, start, len, 0);
2775         if (em_cached && !IS_ERR_OR_NULL(em)) {
2776                 BUG_ON(*em_cached);
2777                 atomic_inc(&em->refs);
2778                 *em_cached = em;
2779         }
2780         return em;
2781 }
2782 /*
2783  * basic readpage implementation.  Locked extent state structs are inserted
2784  * into the tree that are removed when the IO is done (by the end_io
2785  * handlers)
2786  * XXX JDM: This needs looking at to ensure proper page locking
2787  */
2788 static int __do_readpage(struct extent_io_tree *tree,
2789                          struct page *page,
2790                          get_extent_t *get_extent,
2791                          struct extent_map **em_cached,
2792                          struct bio **bio, int mirror_num,
2793                          unsigned long *bio_flags, int rw)
2794 {
2795         struct inode *inode = page->mapping->host;
2796         u64 start = page_offset(page);
2797         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2798         u64 end;
2799         u64 cur = start;
2800         u64 extent_offset;
2801         u64 last_byte = i_size_read(inode);
2802         u64 block_start;
2803         u64 cur_end;
2804         sector_t sector;
2805         struct extent_map *em;
2806         struct block_device *bdev;
2807         int ret;
2808         int nr = 0;
2809         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2810         size_t pg_offset = 0;
2811         size_t iosize;
2812         size_t disk_io_size;
2813         size_t blocksize = inode->i_sb->s_blocksize;
2814         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2815
2816         set_page_extent_mapped(page);
2817
2818         end = page_end;
2819         if (!PageUptodate(page)) {
2820                 if (cleancache_get_page(page) == 0) {
2821                         BUG_ON(blocksize != PAGE_SIZE);
2822                         unlock_extent(tree, start, end);
2823                         goto out;
2824                 }
2825         }
2826
2827         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2828                 char *userpage;
2829                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2830
2831                 if (zero_offset) {
2832                         iosize = PAGE_CACHE_SIZE - zero_offset;
2833                         userpage = kmap_atomic(page);
2834                         memset(userpage + zero_offset, 0, iosize);
2835                         flush_dcache_page(page);
2836                         kunmap_atomic(userpage);
2837                 }
2838         }
2839         while (cur <= end) {
2840                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2841
2842                 if (cur >= last_byte) {
2843                         char *userpage;
2844                         struct extent_state *cached = NULL;
2845
2846                         iosize = PAGE_CACHE_SIZE - pg_offset;
2847                         userpage = kmap_atomic(page);
2848                         memset(userpage + pg_offset, 0, iosize);
2849                         flush_dcache_page(page);
2850                         kunmap_atomic(userpage);
2851                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2852                                             &cached, GFP_NOFS);
2853                         if (!parent_locked)
2854                                 unlock_extent_cached(tree, cur,
2855                                                      cur + iosize - 1,
2856                                                      &cached, GFP_NOFS);
2857                         break;
2858                 }
2859                 em = __get_extent_map(inode, page, pg_offset, cur,
2860                                       end - cur + 1, get_extent, em_cached);
2861                 if (IS_ERR_OR_NULL(em)) {
2862                         SetPageError(page);
2863                         if (!parent_locked)
2864                                 unlock_extent(tree, cur, end);
2865                         break;
2866                 }
2867                 extent_offset = cur - em->start;
2868                 BUG_ON(extent_map_end(em) <= cur);
2869                 BUG_ON(end < cur);
2870
2871                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2872                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2873                         extent_set_compress_type(&this_bio_flag,
2874                                                  em->compress_type);
2875                 }
2876
2877                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2878                 cur_end = min(extent_map_end(em) - 1, end);
2879                 iosize = ALIGN(iosize, blocksize);
2880                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2881                         disk_io_size = em->block_len;
2882                         sector = em->block_start >> 9;
2883                 } else {
2884                         sector = (em->block_start + extent_offset) >> 9;
2885                         disk_io_size = iosize;
2886                 }
2887                 bdev = em->bdev;
2888                 block_start = em->block_start;
2889                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2890                         block_start = EXTENT_MAP_HOLE;
2891                 free_extent_map(em);
2892                 em = NULL;
2893
2894                 /* we've found a hole, just zero and go on */
2895                 if (block_start == EXTENT_MAP_HOLE) {
2896                         char *userpage;
2897                         struct extent_state *cached = NULL;
2898
2899                         userpage = kmap_atomic(page);
2900                         memset(userpage + pg_offset, 0, iosize);
2901                         flush_dcache_page(page);
2902                         kunmap_atomic(userpage);
2903
2904                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2905                                             &cached, GFP_NOFS);
2906                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2907                                              &cached, GFP_NOFS);
2908                         cur = cur + iosize;
2909                         pg_offset += iosize;
2910                         continue;
2911                 }
2912                 /* the get_extent function already copied into the page */
2913                 if (test_range_bit(tree, cur, cur_end,
2914                                    EXTENT_UPTODATE, 1, NULL)) {
2915                         check_page_uptodate(tree, page);
2916                         if (!parent_locked)
2917                                 unlock_extent(tree, cur, cur + iosize - 1);
2918                         cur = cur + iosize;
2919                         pg_offset += iosize;
2920                         continue;
2921                 }
2922                 /* we have an inline extent but it didn't get marked up
2923                  * to date.  Error out
2924                  */
2925                 if (block_start == EXTENT_MAP_INLINE) {
2926                         SetPageError(page);
2927                         if (!parent_locked)
2928                                 unlock_extent(tree, cur, cur + iosize - 1);
2929                         cur = cur + iosize;
2930                         pg_offset += iosize;
2931                         continue;
2932                 }
2933
2934                 pnr -= page->index;
2935                 ret = submit_extent_page(rw, tree, page,
2936                                          sector, disk_io_size, pg_offset,
2937                                          bdev, bio, pnr,
2938                                          end_bio_extent_readpage, mirror_num,
2939                                          *bio_flags,
2940                                          this_bio_flag);
2941                 if (!ret) {
2942                         nr++;
2943                         *bio_flags = this_bio_flag;
2944                 } else {
2945                         SetPageError(page);
2946                         if (!parent_locked)
2947                                 unlock_extent(tree, cur, cur + iosize - 1);
2948                 }
2949                 cur = cur + iosize;
2950                 pg_offset += iosize;
2951         }
2952 out:
2953         if (!nr) {
2954                 if (!PageError(page))
2955                         SetPageUptodate(page);
2956                 unlock_page(page);
2957         }
2958         return 0;
2959 }
2960
2961 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
2962                                              struct page *pages[], int nr_pages,
2963                                              u64 start, u64 end,
2964                                              get_extent_t *get_extent,
2965                                              struct extent_map **em_cached,
2966                                              struct bio **bio, int mirror_num,
2967                                              unsigned long *bio_flags, int rw)
2968 {
2969         struct inode *inode;
2970         struct btrfs_ordered_extent *ordered;
2971         int index;
2972
2973         inode = pages[0]->mapping->host;
2974         while (1) {
2975                 lock_extent(tree, start, end);
2976                 ordered = btrfs_lookup_ordered_range(inode, start,
2977                                                      end - start + 1);
2978                 if (!ordered)
2979                         break;
2980                 unlock_extent(tree, start, end);
2981                 btrfs_start_ordered_extent(inode, ordered, 1);
2982                 btrfs_put_ordered_extent(ordered);
2983         }
2984
2985         for (index = 0; index < nr_pages; index++) {
2986                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
2987                               mirror_num, bio_flags, rw);
2988                 page_cache_release(pages[index]);
2989         }
2990 }
2991
2992 static void __extent_readpages(struct extent_io_tree *tree,
2993                                struct page *pages[],
2994                                int nr_pages, get_extent_t *get_extent,
2995                                struct extent_map **em_cached,
2996                                struct bio **bio, int mirror_num,
2997                                unsigned long *bio_flags, int rw)
2998 {
2999         u64 start = 0;
3000         u64 end = 0;
3001         u64 page_start;
3002         int index;
3003         int first_index = 0;
3004
3005         for (index = 0; index < nr_pages; index++) {
3006                 page_start = page_offset(pages[index]);
3007                 if (!end) {
3008                         start = page_start;
3009                         end = start + PAGE_CACHE_SIZE - 1;
3010                         first_index = index;
3011                 } else if (end + 1 == page_start) {
3012                         end += PAGE_CACHE_SIZE;
3013                 } else {
3014                         __do_contiguous_readpages(tree, &pages[first_index],
3015                                                   index - first_index, start,
3016                                                   end, get_extent, em_cached,
3017                                                   bio, mirror_num, bio_flags,
3018                                                   rw);
3019                         start = page_start;
3020                         end = start + PAGE_CACHE_SIZE - 1;
3021                         first_index = index;
3022                 }
3023         }
3024
3025         if (end)
3026                 __do_contiguous_readpages(tree, &pages[first_index],
3027                                           index - first_index, start,
3028                                           end, get_extent, em_cached, bio,
3029                                           mirror_num, bio_flags, rw);
3030 }
3031
3032 static int __extent_read_full_page(struct extent_io_tree *tree,
3033                                    struct page *page,
3034                                    get_extent_t *get_extent,
3035                                    struct bio **bio, int mirror_num,
3036                                    unsigned long *bio_flags, int rw)
3037 {
3038         struct inode *inode = page->mapping->host;
3039         struct btrfs_ordered_extent *ordered;
3040         u64 start = page_offset(page);
3041         u64 end = start + PAGE_CACHE_SIZE - 1;
3042         int ret;
3043
3044         while (1) {
3045                 lock_extent(tree, start, end);
3046                 ordered = btrfs_lookup_ordered_extent(inode, start);
3047                 if (!ordered)
3048                         break;
3049                 unlock_extent(tree, start, end);
3050                 btrfs_start_ordered_extent(inode, ordered, 1);
3051                 btrfs_put_ordered_extent(ordered);
3052         }
3053
3054         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3055                             bio_flags, rw);
3056         return ret;
3057 }
3058
3059 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3060                             get_extent_t *get_extent, int mirror_num)
3061 {
3062         struct bio *bio = NULL;
3063         unsigned long bio_flags = 0;
3064         int ret;
3065
3066         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3067                                       &bio_flags, READ);
3068         if (bio)
3069                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3070         return ret;
3071 }
3072
3073 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3074                                  get_extent_t *get_extent, int mirror_num)
3075 {
3076         struct bio *bio = NULL;
3077         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3078         int ret;
3079
3080         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3081                                       &bio_flags, READ);
3082         if (bio)
3083                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3084         return ret;
3085 }
3086
3087 static noinline void update_nr_written(struct page *page,
3088                                       struct writeback_control *wbc,
3089                                       unsigned long nr_written)
3090 {
3091         wbc->nr_to_write -= nr_written;
3092         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3093             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3094                 page->mapping->writeback_index = page->index + nr_written;
3095 }
3096
3097 /*
3098  * the writepage semantics are similar to regular writepage.  extent
3099  * records are inserted to lock ranges in the tree, and as dirty areas
3100  * are found, they are marked writeback.  Then the lock bits are removed
3101  * and the end_io handler clears the writeback ranges
3102  */
3103 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3104                               void *data)
3105 {
3106         struct inode *inode = page->mapping->host;
3107         struct extent_page_data *epd = data;
3108         struct extent_io_tree *tree = epd->tree;
3109         u64 start = page_offset(page);
3110         u64 delalloc_start;
3111         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3112         u64 end;
3113         u64 cur = start;
3114         u64 extent_offset;
3115         u64 last_byte = i_size_read(inode);
3116         u64 block_start;
3117         u64 iosize;
3118         sector_t sector;
3119         struct extent_state *cached_state = NULL;
3120         struct extent_map *em;
3121         struct block_device *bdev;
3122         int ret;
3123         int nr = 0;
3124         size_t pg_offset = 0;
3125         size_t blocksize;
3126         loff_t i_size = i_size_read(inode);
3127         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3128         u64 nr_delalloc;
3129         u64 delalloc_end;
3130         int page_started;
3131         int compressed;
3132         int write_flags;
3133         unsigned long nr_written = 0;
3134         bool fill_delalloc = true;
3135
3136         if (wbc->sync_mode == WB_SYNC_ALL)
3137                 write_flags = WRITE_SYNC;
3138         else
3139                 write_flags = WRITE;
3140
3141         trace___extent_writepage(page, inode, wbc);
3142
3143         WARN_ON(!PageLocked(page));
3144
3145         ClearPageError(page);
3146
3147         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3148         if (page->index > end_index ||
3149            (page->index == end_index && !pg_offset)) {
3150                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3151                 unlock_page(page);
3152                 return 0;
3153         }
3154
3155         if (page->index == end_index) {
3156                 char *userpage;
3157
3158                 userpage = kmap_atomic(page);
3159                 memset(userpage + pg_offset, 0,
3160                        PAGE_CACHE_SIZE - pg_offset);
3161                 kunmap_atomic(userpage);
3162                 flush_dcache_page(page);
3163         }
3164         pg_offset = 0;
3165
3166         set_page_extent_mapped(page);
3167
3168         if (!tree->ops || !tree->ops->fill_delalloc)
3169                 fill_delalloc = false;
3170
3171         delalloc_start = start;
3172         delalloc_end = 0;
3173         page_started = 0;
3174         if (!epd->extent_locked && fill_delalloc) {
3175                 u64 delalloc_to_write = 0;
3176                 /*
3177                  * make sure the wbc mapping index is at least updated
3178                  * to this page.
3179                  */
3180                 update_nr_written(page, wbc, 0);
3181
3182                 while (delalloc_end < page_end) {
3183                         nr_delalloc = find_lock_delalloc_range(inode, tree,
3184                                                        page,
3185                                                        &delalloc_start,
3186                                                        &delalloc_end,
3187                                                        128 * 1024 * 1024);
3188                         if (nr_delalloc == 0) {
3189                                 delalloc_start = delalloc_end + 1;
3190                                 continue;
3191                         }
3192                         ret = tree->ops->fill_delalloc(inode, page,
3193                                                        delalloc_start,
3194                                                        delalloc_end,
3195                                                        &page_started,
3196                                                        &nr_written);
3197                         /* File system has been set read-only */
3198                         if (ret) {
3199                                 SetPageError(page);
3200                                 goto done;
3201                         }
3202                         /*
3203                          * delalloc_end is already one less than the total
3204                          * length, so we don't subtract one from
3205                          * PAGE_CACHE_SIZE
3206                          */
3207                         delalloc_to_write += (delalloc_end - delalloc_start +
3208                                               PAGE_CACHE_SIZE) >>
3209                                               PAGE_CACHE_SHIFT;
3210                         delalloc_start = delalloc_end + 1;
3211                 }
3212                 if (wbc->nr_to_write < delalloc_to_write) {
3213                         int thresh = 8192;
3214
3215                         if (delalloc_to_write < thresh * 2)
3216                                 thresh = delalloc_to_write;
3217                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
3218                                                  thresh);
3219                 }
3220
3221                 /* did the fill delalloc function already unlock and start
3222                  * the IO?
3223                  */
3224                 if (page_started) {
3225                         ret = 0;
3226                         /*
3227                          * we've unlocked the page, so we can't update
3228                          * the mapping's writeback index, just update
3229                          * nr_to_write.
3230                          */
3231                         wbc->nr_to_write -= nr_written;
3232                         goto done_unlocked;
3233                 }
3234         }
3235         if (tree->ops && tree->ops->writepage_start_hook) {
3236                 ret = tree->ops->writepage_start_hook(page, start,
3237                                                       page_end);
3238                 if (ret) {
3239                         /* Fixup worker will requeue */
3240                         if (ret == -EBUSY)
3241                                 wbc->pages_skipped++;
3242                         else
3243                                 redirty_page_for_writepage(wbc, page);
3244                         update_nr_written(page, wbc, nr_written);
3245                         unlock_page(page);
3246                         ret = 0;
3247                         goto done_unlocked;
3248                 }
3249         }
3250
3251         /*
3252          * we don't want to touch the inode after unlocking the page,
3253          * so we update the mapping writeback index now
3254          */
3255         update_nr_written(page, wbc, nr_written + 1);
3256
3257         end = page_end;
3258         if (last_byte <= start) {
3259                 if (tree->ops && tree->ops->writepage_end_io_hook)
3260                         tree->ops->writepage_end_io_hook(page, start,
3261                                                          page_end, NULL, 1);
3262                 goto done;
3263         }
3264
3265         blocksize = inode->i_sb->s_blocksize;
3266
3267         while (cur <= end) {
3268                 if (cur >= last_byte) {
3269                         if (tree->ops && tree->ops->writepage_end_io_hook)
3270                                 tree->ops->writepage_end_io_hook(page, cur,
3271                                                          page_end, NULL, 1);
3272                         break;
3273                 }
3274                 em = epd->get_extent(inode, page, pg_offset, cur,
3275                                      end - cur + 1, 1);
3276                 if (IS_ERR_OR_NULL(em)) {
3277                         SetPageError(page);
3278                         break;
3279                 }
3280
3281                 extent_offset = cur - em->start;
3282                 BUG_ON(extent_map_end(em) <= cur);
3283                 BUG_ON(end < cur);
3284                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
3285                 iosize = ALIGN(iosize, blocksize);
3286                 sector = (em->block_start + extent_offset) >> 9;
3287                 bdev = em->bdev;
3288                 block_start = em->block_start;
3289                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3290                 free_extent_map(em);
3291                 em = NULL;
3292
3293                 /*
3294                  * compressed and inline extents are written through other
3295                  * paths in the FS
3296                  */
3297                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3298                     block_start == EXTENT_MAP_INLINE) {
3299                         /*
3300                          * end_io notification does not happen here for
3301                          * compressed extents
3302                          */
3303                         if (!compressed && tree->ops &&
3304                             tree->ops->writepage_end_io_hook)
3305                                 tree->ops->writepage_end_io_hook(page, cur,
3306                                                          cur + iosize - 1,
3307                                                          NULL, 1);
3308                         else if (compressed) {
3309                                 /* we don't want to end_page_writeback on
3310                                  * a compressed extent.  this happens
3311                                  * elsewhere
3312                                  */
3313                                 nr++;
3314                         }
3315
3316                         cur += iosize;
3317                         pg_offset += iosize;
3318                         continue;
3319                 }
3320                 /* leave this out until we have a page_mkwrite call */
3321                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
3322                                    EXTENT_DIRTY, 0, NULL)) {
3323                         cur = cur + iosize;
3324                         pg_offset += iosize;
3325                         continue;
3326                 }
3327
3328                 if (tree->ops && tree->ops->writepage_io_hook) {
3329                         ret = tree->ops->writepage_io_hook(page, cur,
3330                                                 cur + iosize - 1);
3331                 } else {
3332                         ret = 0;
3333                 }
3334                 if (ret) {
3335                         SetPageError(page);
3336                 } else {
3337                         unsigned long max_nr = end_index + 1;
3338
3339                         set_range_writeback(tree, cur, cur + iosize - 1);
3340                         if (!PageWriteback(page)) {
3341                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
3342                                            "page %lu not writeback, cur %llu end %llu",
3343                                        page->index, cur, end);
3344                         }
3345
3346                         ret = submit_extent_page(write_flags, tree, page,
3347                                                  sector, iosize, pg_offset,
3348                                                  bdev, &epd->bio, max_nr,
3349                                                  end_bio_extent_writepage,
3350                                                  0, 0, 0);
3351                         if (ret)
3352                                 SetPageError(page);
3353                 }
3354                 cur = cur + iosize;
3355                 pg_offset += iosize;
3356                 nr++;
3357         }
3358 done:
3359         if (nr == 0) {
3360                 /* make sure the mapping tag for page dirty gets cleared */
3361                 set_page_writeback(page);
3362                 end_page_writeback(page);
3363         }
3364         unlock_page(page);
3365
3366 done_unlocked:
3367
3368         /* drop our reference on any cached states */
3369         free_extent_state(cached_state);
3370         return 0;
3371 }
3372
3373 static int eb_wait(void *word)
3374 {
3375         io_schedule();
3376         return 0;
3377 }
3378
3379 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3380 {
3381         wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
3382                     TASK_UNINTERRUPTIBLE);
3383 }
3384
3385 static int lock_extent_buffer_for_io(struct extent_buffer *eb,
3386                                      struct btrfs_fs_info *fs_info,
3387                                      struct extent_page_data *epd)
3388 {
3389         unsigned long i, num_pages;
3390         int flush = 0;
3391         int ret = 0;
3392
3393         if (!btrfs_try_tree_write_lock(eb)) {
3394                 flush = 1;
3395                 flush_write_bio(epd);
3396                 btrfs_tree_lock(eb);
3397         }
3398
3399         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3400                 btrfs_tree_unlock(eb);
3401                 if (!epd->sync_io)
3402                         return 0;
3403                 if (!flush) {
3404                         flush_write_bio(epd);
3405                         flush = 1;
3406                 }
3407                 while (1) {
3408                         wait_on_extent_buffer_writeback(eb);
3409                         btrfs_tree_lock(eb);
3410                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3411                                 break;
3412                         btrfs_tree_unlock(eb);
3413                 }
3414         }
3415
3416         /*
3417          * We need to do this to prevent races in people who check if the eb is
3418          * under IO since we can end up having no IO bits set for a short period
3419          * of time.
3420          */
3421         spin_lock(&eb->refs_lock);
3422         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3423                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3424                 spin_unlock(&eb->refs_lock);
3425                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3426                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3427                                      -eb->len,
3428                                      fs_info->dirty_metadata_batch);
3429                 ret = 1;
3430         } else {
3431                 spin_unlock(&eb->refs_lock);
3432         }
3433
3434         btrfs_tree_unlock(eb);
3435
3436         if (!ret)
3437                 return ret;
3438
3439         num_pages = num_extent_pages(eb->start, eb->len);
3440         for (i = 0; i < num_pages; i++) {
3441                 struct page *p = extent_buffer_page(eb, i);
3442
3443                 if (!trylock_page(p)) {
3444                         if (!flush) {
3445                                 flush_write_bio(epd);
3446                                 flush = 1;
3447                         }
3448                         lock_page(p);
3449                 }
3450         }
3451
3452         return ret;
3453 }
3454
3455 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3456 {
3457         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3458         smp_mb__after_clear_bit();
3459         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3460 }
3461
3462 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3463 {
3464         struct bio_vec *bvec;
3465         struct extent_buffer *eb;
3466         int i, done;
3467
3468         bio_for_each_segment_all(bvec, bio, i) {
3469                 struct page *page = bvec->bv_page;
3470
3471                 eb = (struct extent_buffer *)page->private;
3472                 BUG_ON(!eb);
3473                 done = atomic_dec_and_test(&eb->io_pages);
3474
3475                 if (err || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3476                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3477                         ClearPageUptodate(page);
3478                         SetPageError(page);
3479                 }
3480
3481                 end_page_writeback(page);
3482
3483                 if (!done)
3484                         continue;
3485
3486                 end_extent_buffer_writeback(eb);
3487         }
3488
3489         bio_put(bio);
3490 }
3491
3492 static int write_one_eb(struct extent_buffer *eb,
3493                         struct btrfs_fs_info *fs_info,
3494                         struct writeback_control *wbc,
3495                         struct extent_page_data *epd)
3496 {
3497         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3498         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3499         u64 offset = eb->start;
3500         unsigned long i, num_pages;
3501         unsigned long bio_flags = 0;
3502         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3503         int ret = 0;
3504
3505         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3506         num_pages = num_extent_pages(eb->start, eb->len);
3507         atomic_set(&eb->io_pages, num_pages);
3508         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3509                 bio_flags = EXTENT_BIO_TREE_LOG;
3510
3511         for (i = 0; i < num_pages; i++) {
3512                 struct page *p = extent_buffer_page(eb, i);
3513
3514                 clear_page_dirty_for_io(p);
3515                 set_page_writeback(p);
3516                 ret = submit_extent_page(rw, tree, p, offset >> 9,
3517                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3518                                          -1, end_bio_extent_buffer_writepage,
3519                                          0, epd->bio_flags, bio_flags);
3520                 epd->bio_flags = bio_flags;
3521                 if (ret) {
3522                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3523                         SetPageError(p);
3524                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3525                                 end_extent_buffer_writeback(eb);
3526                         ret = -EIO;
3527                         break;
3528                 }
3529                 offset += PAGE_CACHE_SIZE;
3530                 update_nr_written(p, wbc, 1);
3531                 unlock_page(p);
3532         }
3533
3534         if (unlikely(ret)) {
3535                 for (; i < num_pages; i++) {
3536                         struct page *p = extent_buffer_page(eb, i);
3537                         unlock_page(p);
3538                 }
3539         }
3540
3541         return ret;
3542 }
3543
3544 int btree_write_cache_pages(struct address_space *mapping,
3545                                    struct writeback_control *wbc)
3546 {
3547         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3548         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3549         struct extent_buffer *eb, *prev_eb = NULL;
3550         struct extent_page_data epd = {
3551                 .bio = NULL,
3552                 .tree = tree,
3553                 .extent_locked = 0,
3554                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3555                 .bio_flags = 0,
3556         };
3557         int ret = 0;
3558         int done = 0;
3559         int nr_to_write_done = 0;
3560         struct pagevec pvec;
3561         int nr_pages;
3562         pgoff_t index;
3563         pgoff_t end;            /* Inclusive */
3564         int scanned = 0;
3565         int tag;
3566
3567         pagevec_init(&pvec, 0);
3568         if (wbc->range_cyclic) {
3569                 index = mapping->writeback_index; /* Start from prev offset */
3570                 end = -1;
3571         } else {
3572                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3573                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3574                 scanned = 1;
3575         }
3576         if (wbc->sync_mode == WB_SYNC_ALL)
3577                 tag = PAGECACHE_TAG_TOWRITE;
3578         else
3579                 tag = PAGECACHE_TAG_DIRTY;
3580 retry:
3581         if (wbc->sync_mode == WB_SYNC_ALL)
3582                 tag_pages_for_writeback(mapping, index, end);
3583         while (!done && !nr_to_write_done && (index <= end) &&
3584                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3585                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3586                 unsigned i;
3587
3588                 scanned = 1;
3589                 for (i = 0; i < nr_pages; i++) {
3590                         struct page *page = pvec.pages[i];
3591
3592                         if (!PagePrivate(page))
3593                                 continue;
3594
3595                         if (!wbc->range_cyclic && page->index > end) {
3596                                 done = 1;
3597                                 break;
3598                         }
3599
3600                         spin_lock(&mapping->private_lock);
3601                         if (!PagePrivate(page)) {
3602                                 spin_unlock(&mapping->private_lock);
3603                                 continue;
3604                         }
3605
3606                         eb = (struct extent_buffer *)page->private;
3607
3608                         /*
3609                          * Shouldn't happen and normally this would be a BUG_ON
3610                          * but no sense in crashing the users box for something
3611                          * we can survive anyway.
3612                          */
3613                         if (WARN_ON(!eb)) {
3614                                 spin_unlock(&mapping->private_lock);
3615                                 continue;
3616                         }
3617
3618                         if (eb == prev_eb) {
3619                                 spin_unlock(&mapping->private_lock);
3620                                 continue;
3621                         }
3622
3623                         ret = atomic_inc_not_zero(&eb->refs);
3624                         spin_unlock(&mapping->private_lock);
3625                         if (!ret)
3626                                 continue;
3627
3628                         prev_eb = eb;
3629                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3630                         if (!ret) {
3631                                 free_extent_buffer(eb);
3632                                 continue;
3633                         }
3634
3635                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3636                         if (ret) {
3637                                 done = 1;
3638                                 free_extent_buffer(eb);
3639                                 break;
3640                         }
3641                         free_extent_buffer(eb);
3642
3643                         /*
3644                          * the filesystem may choose to bump up nr_to_write.
3645                          * We have to make sure to honor the new nr_to_write
3646                          * at any time
3647                          */
3648                         nr_to_write_done = wbc->nr_to_write <= 0;
3649                 }
3650                 pagevec_release(&pvec);
3651                 cond_resched();
3652         }
3653         if (!scanned && !done) {
3654                 /*
3655                  * We hit the last page and there is more work to be done: wrap
3656                  * back to the start of the file
3657                  */
3658                 scanned = 1;
3659                 index = 0;
3660                 goto retry;
3661         }
3662         flush_write_bio(&epd);
3663         return ret;
3664 }
3665
3666 /**
3667  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3668  * @mapping: address space structure to write
3669  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3670  * @writepage: function called for each page
3671  * @data: data passed to writepage function
3672  *
3673  * If a page is already under I/O, write_cache_pages() skips it, even
3674  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3675  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3676  * and msync() need to guarantee that all the data which was dirty at the time
3677  * the call was made get new I/O started against them.  If wbc->sync_mode is
3678  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3679  * existing IO to complete.
3680  */
3681 static int extent_write_cache_pages(struct extent_io_tree *tree,
3682                              struct address_space *mapping,
3683                              struct writeback_control *wbc,
3684                              writepage_t writepage, void *data,
3685                              void (*flush_fn)(void *))
3686 {
3687         struct inode *inode = mapping->host;
3688         int ret = 0;
3689         int done = 0;
3690         int nr_to_write_done = 0;
3691         struct pagevec pvec;
3692         int nr_pages;
3693         pgoff_t index;
3694         pgoff_t end;            /* Inclusive */
3695         int scanned = 0;
3696         int tag;
3697
3698         /*
3699          * We have to hold onto the inode so that ordered extents can do their
3700          * work when the IO finishes.  The alternative to this is failing to add
3701          * an ordered extent if the igrab() fails there and that is a huge pain
3702          * to deal with, so instead just hold onto the inode throughout the
3703          * writepages operation.  If it fails here we are freeing up the inode
3704          * anyway and we'd rather not waste our time writing out stuff that is
3705          * going to be truncated anyway.
3706          */
3707         if (!igrab(inode))
3708                 return 0;
3709
3710         pagevec_init(&pvec, 0);
3711         if (wbc->range_cyclic) {
3712                 index = mapping->writeback_index; /* Start from prev offset */
3713                 end = -1;
3714         } else {
3715                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3716                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3717                 scanned = 1;
3718         }
3719         if (wbc->sync_mode == WB_SYNC_ALL)
3720                 tag = PAGECACHE_TAG_TOWRITE;
3721         else
3722                 tag = PAGECACHE_TAG_DIRTY;
3723 retry:
3724         if (wbc->sync_mode == WB_SYNC_ALL)
3725                 tag_pages_for_writeback(mapping, index, end);
3726         while (!done && !nr_to_write_done && (index <= end) &&
3727                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3728                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3729                 unsigned i;
3730
3731                 scanned = 1;
3732                 for (i = 0; i < nr_pages; i++) {
3733                         struct page *page = pvec.pages[i];
3734
3735                         /*
3736                          * At this point we hold neither mapping->tree_lock nor
3737                          * lock on the page itself: the page may be truncated or
3738                          * invalidated (changing page->mapping to NULL), or even
3739                          * swizzled back from swapper_space to tmpfs file
3740                          * mapping
3741                          */
3742                         if (!trylock_page(page)) {
3743                                 flush_fn(data);
3744                                 lock_page(page);
3745                         }
3746
3747                         if (unlikely(page->mapping != mapping)) {
3748                                 unlock_page(page);
3749                                 continue;
3750                         }
3751
3752                         if (!wbc->range_cyclic && page->index > end) {
3753                                 done = 1;
3754                                 unlock_page(page);
3755                                 continue;
3756                         }
3757
3758                         if (wbc->sync_mode != WB_SYNC_NONE) {
3759                                 if (PageWriteback(page))
3760                                         flush_fn(data);
3761                                 wait_on_page_writeback(page);
3762                         }
3763
3764                         if (PageWriteback(page) ||
3765                             !clear_page_dirty_for_io(page)) {
3766                                 unlock_page(page);
3767                                 continue;
3768                         }
3769
3770                         ret = (*writepage)(page, wbc, data);
3771
3772                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3773                                 unlock_page(page);
3774                                 ret = 0;
3775                         }
3776                         if (ret)
3777                                 done = 1;
3778
3779                         /*
3780                          * the filesystem may choose to bump up nr_to_write.
3781                          * We have to make sure to honor the new nr_to_write
3782                          * at any time
3783                          */
3784                         nr_to_write_done = wbc->nr_to_write <= 0;
3785                 }
3786                 pagevec_release(&pvec);
3787                 cond_resched();
3788         }
3789         if (!scanned && !done) {
3790                 /*
3791                  * We hit the last page and there is more work to be done: wrap
3792                  * back to the start of the file
3793                  */
3794                 scanned = 1;
3795                 index = 0;
3796                 goto retry;
3797         }
3798         btrfs_add_delayed_iput(inode);
3799         return ret;
3800 }
3801
3802 static void flush_epd_write_bio(struct extent_page_data *epd)
3803 {
3804         if (epd->bio) {
3805                 int rw = WRITE;
3806                 int ret;
3807
3808                 if (epd->sync_io)
3809                         rw = WRITE_SYNC;
3810
3811                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
3812                 BUG_ON(ret < 0); /* -ENOMEM */
3813                 epd->bio = NULL;
3814         }
3815 }
3816
3817 static noinline void flush_write_bio(void *data)
3818 {
3819         struct extent_page_data *epd = data;
3820         flush_epd_write_bio(epd);
3821 }
3822
3823 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3824                           get_extent_t *get_extent,
3825                           struct writeback_control *wbc)
3826 {
3827         int ret;
3828         struct extent_page_data epd = {
3829                 .bio = NULL,
3830                 .tree = tree,
3831                 .get_extent = get_extent,
3832                 .extent_locked = 0,
3833                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3834                 .bio_flags = 0,
3835         };
3836
3837         ret = __extent_writepage(page, wbc, &epd);
3838
3839         flush_epd_write_bio(&epd);
3840         return ret;
3841 }
3842
3843 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3844                               u64 start, u64 end, get_extent_t *get_extent,
3845                               int mode)
3846 {
3847         int ret = 0;
3848         struct address_space *mapping = inode->i_mapping;
3849         struct page *page;
3850         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3851                 PAGE_CACHE_SHIFT;
3852
3853         struct extent_page_data epd = {
3854                 .bio = NULL,
3855                 .tree = tree,
3856                 .get_extent = get_extent,
3857                 .extent_locked = 1,
3858                 .sync_io = mode == WB_SYNC_ALL,
3859                 .bio_flags = 0,
3860         };
3861         struct writeback_control wbc_writepages = {
3862                 .sync_mode      = mode,
3863                 .nr_to_write    = nr_pages * 2,
3864                 .range_start    = start,
3865                 .range_end      = end + 1,
3866         };
3867
3868         while (start <= end) {
3869                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3870                 if (clear_page_dirty_for_io(page))
3871                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3872                 else {
3873                         if (tree->ops && tree->ops->writepage_end_io_hook)
3874                                 tree->ops->writepage_end_io_hook(page, start,
3875                                                  start + PAGE_CACHE_SIZE - 1,
3876                                                  NULL, 1);
3877                         unlock_page(page);
3878                 }
3879                 page_cache_release(page);
3880                 start += PAGE_CACHE_SIZE;
3881         }
3882
3883         flush_epd_write_bio(&epd);
3884         return ret;
3885 }
3886
3887 int extent_writepages(struct extent_io_tree *tree,
3888                       struct address_space *mapping,
3889                       get_extent_t *get_extent,
3890                       struct writeback_control *wbc)
3891 {
3892         int ret = 0;
3893         struct extent_page_data epd = {
3894                 .bio = NULL,
3895                 .tree = tree,
3896                 .get_extent = get_extent,
3897                 .extent_locked = 0,
3898                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3899                 .bio_flags = 0,
3900         };
3901
3902         ret = extent_write_cache_pages(tree, mapping, wbc,
3903                                        __extent_writepage, &epd,
3904                                        flush_write_bio);
3905         flush_epd_write_bio(&epd);
3906         return ret;
3907 }
3908
3909 int extent_readpages(struct extent_io_tree *tree,
3910                      struct address_space *mapping,
3911                      struct list_head *pages, unsigned nr_pages,
3912                      get_extent_t get_extent)
3913 {
3914         struct bio *bio = NULL;
3915         unsigned page_idx;
3916         unsigned long bio_flags = 0;
3917         struct page *pagepool[16];
3918         struct page *page;
3919         struct extent_map *em_cached = NULL;
3920         int nr = 0;
3921
3922         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3923                 page = list_entry(pages->prev, struct page, lru);
3924
3925                 prefetchw(&page->flags);
3926                 list_del(&page->lru);
3927                 if (add_to_page_cache_lru(page, mapping,
3928                                         page->index, GFP_NOFS)) {
3929                         page_cache_release(page);
3930                         continue;
3931                 }
3932
3933                 pagepool[nr++] = page;
3934                 if (nr < ARRAY_SIZE(pagepool))
3935                         continue;
3936                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
3937                                    &bio, 0, &bio_flags, READ);
3938                 nr = 0;
3939         }
3940         if (nr)
3941                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
3942                                    &bio, 0, &bio_flags, READ);
3943
3944         if (em_cached)
3945                 free_extent_map(em_cached);
3946
3947         BUG_ON(!list_empty(pages));
3948         if (bio)
3949                 return submit_one_bio(READ, bio, 0, bio_flags);
3950         return 0;
3951 }
3952
3953 /*
3954  * basic invalidatepage code, this waits on any locked or writeback
3955  * ranges corresponding to the page, and then deletes any extent state
3956  * records from the tree
3957  */
3958 int extent_invalidatepage(struct extent_io_tree *tree,
3959                           struct page *page, unsigned long offset)
3960 {
3961         struct extent_state *cached_state = NULL;
3962         u64 start = page_offset(page);
3963         u64 end = start + PAGE_CACHE_SIZE - 1;
3964         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3965
3966         start += ALIGN(offset, blocksize);
3967         if (start > end)
3968                 return 0;
3969
3970         lock_extent_bits(tree, start, end, 0, &cached_state);
3971         wait_on_page_writeback(page);
3972         clear_extent_bit(tree, start, end,
3973                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3974                          EXTENT_DO_ACCOUNTING,
3975                          1, 1, &cached_state, GFP_NOFS);
3976         return 0;
3977 }
3978
3979 /*
3980  * a helper for releasepage, this tests for areas of the page that
3981  * are locked or under IO and drops the related state bits if it is safe
3982  * to drop the page.
3983  */
3984 static int try_release_extent_state(struct extent_map_tree *map,
3985                                     struct extent_io_tree *tree,
3986                                     struct page *page, gfp_t mask)
3987 {
3988         u64 start = page_offset(page);
3989         u64 end = start + PAGE_CACHE_SIZE - 1;
3990         int ret = 1;
3991
3992         if (test_range_bit(tree, start, end,
3993                            EXTENT_IOBITS, 0, NULL))
3994                 ret = 0;
3995         else {
3996                 if ((mask & GFP_NOFS) == GFP_NOFS)
3997                         mask = GFP_NOFS;
3998                 /*
3999                  * at this point we can safely clear everything except the
4000                  * locked bit and the nodatasum bit
4001                  */
4002                 ret = clear_extent_bit(tree, start, end,
4003                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4004                                  0, 0, NULL, mask);
4005
4006                 /* if clear_extent_bit failed for enomem reasons,
4007                  * we can't allow the release to continue.
4008                  */
4009                 if (ret < 0)
4010                         ret = 0;
4011                 else
4012                         ret = 1;
4013         }
4014         return ret;
4015 }
4016
4017 /*
4018  * a helper for releasepage.  As long as there are no locked extents
4019  * in the range corresponding to the page, both state records and extent
4020  * map records are removed
4021  */
4022 int try_release_extent_mapping(struct extent_map_tree *map,
4023                                struct extent_io_tree *tree, struct page *page,
4024                                gfp_t mask)
4025 {
4026         struct extent_map *em;
4027         u64 start = page_offset(page);
4028         u64 end = start + PAGE_CACHE_SIZE - 1;
4029
4030         if ((mask & __GFP_WAIT) &&
4031             page->mapping->host->i_size > 16 * 1024 * 1024) {
4032                 u64 len;
4033                 while (start <= end) {
4034                         len = end - start + 1;
4035                         write_lock(&map->lock);
4036                         em = lookup_extent_mapping(map, start, len);
4037                         if (!em) {
4038                                 write_unlock(&map->lock);
4039                                 break;
4040                         }
4041                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4042                             em->start != start) {
4043                                 write_unlock(&map->lock);
4044                                 free_extent_map(em);
4045                                 break;
4046                         }
4047                         if (!test_range_bit(tree, em->start,
4048                                             extent_map_end(em) - 1,
4049                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4050                                             0, NULL)) {
4051                                 remove_extent_mapping(map, em);
4052                                 /* once for the rb tree */
4053                                 free_extent_map(em);
4054                         }
4055                         start = extent_map_end(em);
4056                         write_unlock(&map->lock);
4057
4058                         /* once for us */
4059                         free_extent_map(em);
4060                 }
4061         }
4062         return try_release_extent_state(map, tree, page, mask);
4063 }
4064
4065 /*
4066  * helper function for fiemap, which doesn't want to see any holes.
4067  * This maps until we find something past 'last'
4068  */
4069 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4070                                                 u64 offset,
4071                                                 u64 last,
4072                                                 get_extent_t *get_extent)
4073 {
4074         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4075         struct extent_map *em;
4076         u64 len;
4077
4078         if (offset >= last)
4079                 return NULL;
4080
4081         while (1) {
4082                 len = last - offset;
4083                 if (len == 0)
4084                         break;
4085                 len = ALIGN(len, sectorsize);
4086                 em = get_extent(inode, NULL, 0, offset, len, 0);
4087                 if (IS_ERR_OR_NULL(em))
4088                         return em;
4089
4090                 /* if this isn't a hole return it */
4091                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4092                     em->block_start != EXTENT_MAP_HOLE) {
4093                         return em;
4094                 }
4095
4096                 /* this is a hole, advance to the next extent */
4097                 offset = extent_map_end(em);
4098                 free_extent_map(em);
4099                 if (offset >= last)
4100                         break;
4101         }
4102         return NULL;
4103 }
4104
4105 static noinline int count_ext_ref(u64 inum, u64 offset, u64 root_id, void *ctx)
4106 {
4107         unsigned long cnt = *((unsigned long *)ctx);
4108
4109         cnt++;
4110         *((unsigned long *)ctx) = cnt;
4111
4112         /* Now we're sure that the extent is shared. */
4113         if (cnt > 1)
4114                 return 1;
4115         return 0;
4116 }
4117
4118 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4119                 __u64 start, __u64 len, get_extent_t *get_extent)
4120 {
4121         int ret = 0;
4122         u64 off = start;
4123         u64 max = start + len;
4124         u32 flags = 0;
4125         u32 found_type;
4126         u64 last;
4127         u64 last_for_get_extent = 0;
4128         u64 disko = 0;
4129         u64 isize = i_size_read(inode);
4130         struct btrfs_key found_key;
4131         struct extent_map *em = NULL;
4132         struct extent_state *cached_state = NULL;
4133         struct btrfs_path *path;
4134         int end = 0;
4135         u64 em_start = 0;
4136         u64 em_len = 0;
4137         u64 em_end = 0;
4138
4139         if (len == 0)
4140                 return -EINVAL;
4141
4142         path = btrfs_alloc_path();
4143         if (!path)
4144                 return -ENOMEM;
4145         path->leave_spinning = 1;
4146
4147         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
4148         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
4149
4150         /*
4151          * lookup the last file extent.  We're not using i_size here
4152          * because there might be preallocation past i_size
4153          */
4154         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
4155                                        path, btrfs_ino(inode), -1, 0);
4156         if (ret < 0) {
4157                 btrfs_free_path(path);
4158                 return ret;
4159         }
4160         WARN_ON(!ret);
4161         path->slots[0]--;
4162         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4163         found_type = btrfs_key_type(&found_key);
4164
4165         /* No extents, but there might be delalloc bits */
4166         if (found_key.objectid != btrfs_ino(inode) ||
4167             found_type != BTRFS_EXTENT_DATA_KEY) {
4168                 /* have to trust i_size as the end */
4169                 last = (u64)-1;
4170                 last_for_get_extent = isize;
4171         } else {
4172                 /*
4173                  * remember the start of the last extent.  There are a
4174                  * bunch of different factors that go into the length of the
4175                  * extent, so its much less complex to remember where it started
4176                  */
4177                 last = found_key.offset;
4178                 last_for_get_extent = last + 1;
4179         }
4180         btrfs_release_path(path);
4181
4182         /*
4183          * we might have some extents allocated but more delalloc past those
4184          * extents.  so, we trust isize unless the start of the last extent is
4185          * beyond isize
4186          */
4187         if (last < isize) {
4188                 last = (u64)-1;
4189                 last_for_get_extent = isize;
4190         }
4191
4192         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4193                          &cached_state);
4194
4195         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4196                                    get_extent);
4197         if (!em)
4198                 goto out;
4199         if (IS_ERR(em)) {
4200                 ret = PTR_ERR(em);
4201                 goto out;
4202         }
4203
4204         while (!end) {
4205                 u64 offset_in_extent = 0;
4206
4207                 /* break if the extent we found is outside the range */
4208                 if (em->start >= max || extent_map_end(em) < off)
4209                         break;
4210
4211                 /*
4212                  * get_extent may return an extent that starts before our
4213                  * requested range.  We have to make sure the ranges
4214                  * we return to fiemap always move forward and don't
4215                  * overlap, so adjust the offsets here
4216                  */
4217                 em_start = max(em->start, off);
4218
4219                 /*
4220                  * record the offset from the start of the extent
4221                  * for adjusting the disk offset below.  Only do this if the
4222                  * extent isn't compressed since our in ram offset may be past
4223                  * what we have actually allocated on disk.
4224                  */
4225                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4226                         offset_in_extent = em_start - em->start;
4227                 em_end = extent_map_end(em);
4228                 em_len = em_end - em_start;
4229                 disko = 0;
4230                 flags = 0;
4231
4232                 /*
4233                  * bump off for our next call to get_extent
4234                  */
4235                 off = extent_map_end(em);
4236                 if (off >= max)
4237                         end = 1;
4238
4239                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4240                         end = 1;
4241                         flags |= FIEMAP_EXTENT_LAST;
4242                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4243                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4244                                   FIEMAP_EXTENT_NOT_ALIGNED);
4245                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4246                         flags |= (FIEMAP_EXTENT_DELALLOC |
4247                                   FIEMAP_EXTENT_UNKNOWN);
4248                 } else {
4249                         unsigned long ref_cnt = 0;
4250
4251                         disko = em->block_start + offset_in_extent;
4252
4253                         /*
4254                          * As btrfs supports shared space, this information
4255                          * can be exported to userspace tools via
4256                          * flag FIEMAP_EXTENT_SHARED.
4257                          */
4258                         ret = iterate_inodes_from_logical(
4259                                         em->block_start,
4260                                         BTRFS_I(inode)->root->fs_info,
4261                                         path, count_ext_ref, &ref_cnt);
4262                         if (ret < 0 && ret != -ENOENT)
4263                                 goto out_free;
4264
4265                         if (ref_cnt > 1)
4266                                 flags |= FIEMAP_EXTENT_SHARED;
4267                 }
4268                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4269                         flags |= FIEMAP_EXTENT_ENCODED;
4270
4271                 free_extent_map(em);
4272                 em = NULL;
4273                 if ((em_start >= last) || em_len == (u64)-1 ||
4274                    (last == (u64)-1 && isize <= em_end)) {
4275                         flags |= FIEMAP_EXTENT_LAST;
4276                         end = 1;
4277                 }
4278
4279                 /* now scan forward to see if this is really the last extent. */
4280                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4281                                            get_extent);
4282                 if (IS_ERR(em)) {
4283                         ret = PTR_ERR(em);
4284                         goto out;
4285                 }
4286                 if (!em) {
4287                         flags |= FIEMAP_EXTENT_LAST;
4288                         end = 1;
4289                 }
4290                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4291                                               em_len, flags);
4292                 if (ret)
4293                         goto out_free;
4294         }
4295 out_free:
4296         free_extent_map(em);
4297 out:
4298         btrfs_free_path(path);
4299         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4300                              &cached_state, GFP_NOFS);
4301         return ret;
4302 }
4303
4304 static void __free_extent_buffer(struct extent_buffer *eb)
4305 {
4306         btrfs_leak_debug_del(&eb->leak_list);
4307         kmem_cache_free(extent_buffer_cache, eb);
4308 }
4309
4310 static int extent_buffer_under_io(struct extent_buffer *eb)
4311 {
4312         return (atomic_read(&eb->io_pages) ||
4313                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4314                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4315 }
4316
4317 /*
4318  * Helper for releasing extent buffer page.
4319  */
4320 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
4321                                                 unsigned long start_idx)
4322 {
4323         unsigned long index;
4324         unsigned long num_pages;
4325         struct page *page;
4326         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4327
4328         BUG_ON(extent_buffer_under_io(eb));
4329
4330         num_pages = num_extent_pages(eb->start, eb->len);
4331         index = start_idx + num_pages;
4332         if (start_idx >= index)
4333                 return;
4334
4335         do {
4336                 index--;
4337                 page = extent_buffer_page(eb, index);
4338                 if (page && mapped) {
4339                         spin_lock(&page->mapping->private_lock);
4340                         /*
4341                          * We do this since we'll remove the pages after we've
4342                          * removed the eb from the radix tree, so we could race
4343                          * and have this page now attached to the new eb.  So
4344                          * only clear page_private if it's still connected to
4345                          * this eb.
4346                          */
4347                         if (PagePrivate(page) &&
4348                             page->private == (unsigned long)eb) {
4349                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4350                                 BUG_ON(PageDirty(page));
4351                                 BUG_ON(PageWriteback(page));
4352                                 /*
4353                                  * We need to make sure we haven't be attached
4354                                  * to a new eb.
4355                                  */
4356                                 ClearPagePrivate(page);
4357                                 set_page_private(page, 0);
4358                                 /* One for the page private */
4359                                 page_cache_release(page);
4360                         }
4361                         spin_unlock(&page->mapping->private_lock);
4362
4363                 }
4364                 if (page) {
4365                         /* One for when we alloced the page */
4366                         page_cache_release(page);
4367                 }
4368         } while (index != start_idx);
4369 }
4370
4371 /*
4372  * Helper for releasing the extent buffer.
4373  */
4374 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4375 {
4376         btrfs_release_extent_buffer_page(eb, 0);
4377         __free_extent_buffer(eb);
4378 }
4379
4380 static struct extent_buffer *
4381 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4382                       unsigned long len, gfp_t mask)
4383 {
4384         struct extent_buffer *eb = NULL;
4385
4386         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
4387         if (eb == NULL)
4388                 return NULL;
4389         eb->start = start;
4390         eb->len = len;
4391         eb->fs_info = fs_info;
4392         eb->bflags = 0;
4393         rwlock_init(&eb->lock);
4394         atomic_set(&eb->write_locks, 0);
4395         atomic_set(&eb->read_locks, 0);
4396         atomic_set(&eb->blocking_readers, 0);
4397         atomic_set(&eb->blocking_writers, 0);
4398         atomic_set(&eb->spinning_readers, 0);
4399         atomic_set(&eb->spinning_writers, 0);
4400         eb->lock_nested = 0;
4401         init_waitqueue_head(&eb->write_lock_wq);
4402         init_waitqueue_head(&eb->read_lock_wq);
4403
4404         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4405
4406         spin_lock_init(&eb->refs_lock);
4407         atomic_set(&eb->refs, 1);
4408         atomic_set(&eb->io_pages, 0);
4409
4410         /*
4411          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4412          */
4413         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4414                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4415         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4416
4417         return eb;
4418 }
4419
4420 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4421 {
4422         unsigned long i;
4423         struct page *p;
4424         struct extent_buffer *new;
4425         unsigned long num_pages = num_extent_pages(src->start, src->len);
4426
4427         new = __alloc_extent_buffer(NULL, src->start, src->len, GFP_NOFS);
4428         if (new == NULL)
4429                 return NULL;
4430
4431         for (i = 0; i < num_pages; i++) {
4432                 p = alloc_page(GFP_NOFS);
4433                 if (!p) {
4434                         btrfs_release_extent_buffer(new);
4435                         return NULL;
4436                 }
4437                 attach_extent_buffer_page(new, p);
4438                 WARN_ON(PageDirty(p));
4439                 SetPageUptodate(p);
4440                 new->pages[i] = p;
4441         }
4442
4443         copy_extent_buffer(new, src, 0, 0, src->len);
4444         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4445         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4446
4447         return new;
4448 }
4449
4450 struct extent_buffer *alloc_dummy_extent_buffer(u64 start, unsigned long len)
4451 {
4452         struct extent_buffer *eb;
4453         unsigned long num_pages = num_extent_pages(0, len);
4454         unsigned long i;
4455
4456         eb = __alloc_extent_buffer(NULL, start, len, GFP_NOFS);
4457         if (!eb)
4458                 return NULL;
4459
4460         for (i = 0; i < num_pages; i++) {
4461                 eb->pages[i] = alloc_page(GFP_NOFS);
4462                 if (!eb->pages[i])
4463                         goto err;
4464         }
4465         set_extent_buffer_uptodate(eb);
4466         btrfs_set_header_nritems(eb, 0);
4467         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4468
4469         return eb;
4470 err:
4471         for (; i > 0; i--)
4472                 __free_page(eb->pages[i - 1]);
4473         __free_extent_buffer(eb);
4474         return NULL;
4475 }
4476
4477 static void check_buffer_tree_ref(struct extent_buffer *eb)
4478 {
4479         int refs;
4480         /* the ref bit is tricky.  We have to make sure it is set
4481          * if we have the buffer dirty.   Otherwise the
4482          * code to free a buffer can end up dropping a dirty
4483          * page
4484          *
4485          * Once the ref bit is set, it won't go away while the
4486          * buffer is dirty or in writeback, and it also won't
4487          * go away while we have the reference count on the
4488          * eb bumped.
4489          *
4490          * We can't just set the ref bit without bumping the
4491          * ref on the eb because free_extent_buffer might
4492          * see the ref bit and try to clear it.  If this happens
4493          * free_extent_buffer might end up dropping our original
4494          * ref by mistake and freeing the page before we are able
4495          * to add one more ref.
4496          *
4497          * So bump the ref count first, then set the bit.  If someone
4498          * beat us to it, drop the ref we added.
4499          */
4500         refs = atomic_read(&eb->refs);
4501         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4502                 return;
4503
4504         spin_lock(&eb->refs_lock);
4505         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4506                 atomic_inc(&eb->refs);
4507         spin_unlock(&eb->refs_lock);
4508 }
4509
4510 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4511                 struct page *accessed)
4512 {
4513         unsigned long num_pages, i;
4514
4515         check_buffer_tree_ref(eb);
4516
4517         num_pages = num_extent_pages(eb->start, eb->len);
4518         for (i = 0; i < num_pages; i++) {
4519                 struct page *p = extent_buffer_page(eb, i);
4520                 if (p != accessed)
4521                         mark_page_accessed(p);
4522         }
4523 }
4524
4525 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4526                                          u64 start)
4527 {
4528         struct extent_buffer *eb;
4529
4530         rcu_read_lock();
4531         eb = radix_tree_lookup(&fs_info->buffer_radix,
4532                                start >> PAGE_CACHE_SHIFT);
4533         if (eb && atomic_inc_not_zero(&eb->refs)) {
4534                 rcu_read_unlock();
4535                 mark_extent_buffer_accessed(eb, NULL);
4536                 return eb;
4537         }
4538         rcu_read_unlock();
4539
4540         return NULL;
4541 }
4542
4543 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4544                                           u64 start, unsigned long len)
4545 {
4546         unsigned long num_pages = num_extent_pages(start, len);
4547         unsigned long i;
4548         unsigned long index = start >> PAGE_CACHE_SHIFT;
4549         struct extent_buffer *eb;
4550         struct extent_buffer *exists = NULL;
4551         struct page *p;
4552         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4553         int uptodate = 1;
4554         int ret;
4555
4556         eb = find_extent_buffer(fs_info, start);
4557         if (eb)
4558                 return eb;
4559
4560         eb = __alloc_extent_buffer(fs_info, start, len, GFP_NOFS);
4561         if (!eb)
4562                 return NULL;
4563
4564         for (i = 0; i < num_pages; i++, index++) {
4565                 p = find_or_create_page(mapping, index, GFP_NOFS);
4566                 if (!p)
4567                         goto free_eb;
4568
4569                 spin_lock(&mapping->private_lock);
4570                 if (PagePrivate(p)) {
4571                         /*
4572                          * We could have already allocated an eb for this page
4573                          * and attached one so lets see if we can get a ref on
4574                          * the existing eb, and if we can we know it's good and
4575                          * we can just return that one, else we know we can just
4576                          * overwrite page->private.
4577                          */
4578                         exists = (struct extent_buffer *)p->private;
4579                         if (atomic_inc_not_zero(&exists->refs)) {
4580                                 spin_unlock(&mapping->private_lock);
4581                                 unlock_page(p);
4582                                 page_cache_release(p);
4583                                 mark_extent_buffer_accessed(exists, p);
4584                                 goto free_eb;
4585                         }
4586
4587                         /*
4588                          * Do this so attach doesn't complain and we need to
4589                          * drop the ref the old guy had.
4590                          */
4591                         ClearPagePrivate(p);
4592                         WARN_ON(PageDirty(p));
4593                         page_cache_release(p);
4594                 }
4595                 attach_extent_buffer_page(eb, p);
4596                 spin_unlock(&mapping->private_lock);
4597                 WARN_ON(PageDirty(p));
4598                 eb->pages[i] = p;
4599                 if (!PageUptodate(p))
4600                         uptodate = 0;
4601
4602                 /*
4603                  * see below about how we avoid a nasty race with release page
4604                  * and why we unlock later
4605                  */
4606         }
4607         if (uptodate)
4608                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4609 again:
4610         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4611         if (ret)
4612                 goto free_eb;
4613
4614         spin_lock(&fs_info->buffer_lock);
4615         ret = radix_tree_insert(&fs_info->buffer_radix,
4616                                 start >> PAGE_CACHE_SHIFT, eb);
4617         spin_unlock(&fs_info->buffer_lock);
4618         radix_tree_preload_end();
4619         if (ret == -EEXIST) {
4620                 exists = find_extent_buffer(fs_info, start);
4621                 if (exists)
4622                         goto free_eb;
4623                 else
4624                         goto again;
4625         }
4626         /* add one reference for the tree */
4627         check_buffer_tree_ref(eb);
4628         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4629
4630         /*
4631          * there is a race where release page may have
4632          * tried to find this extent buffer in the radix
4633          * but failed.  It will tell the VM it is safe to
4634          * reclaim the, and it will clear the page private bit.
4635          * We must make sure to set the page private bit properly
4636          * after the extent buffer is in the radix tree so
4637          * it doesn't get lost
4638          */
4639         SetPageChecked(eb->pages[0]);
4640         for (i = 1; i < num_pages; i++) {
4641                 p = extent_buffer_page(eb, i);
4642                 ClearPageChecked(p);
4643                 unlock_page(p);
4644         }
4645         unlock_page(eb->pages[0]);
4646         return eb;
4647
4648 free_eb:
4649         for (i = 0; i < num_pages; i++) {
4650                 if (eb->pages[i])
4651                         unlock_page(eb->pages[i]);
4652         }
4653
4654         WARN_ON(!atomic_dec_and_test(&eb->refs));
4655         btrfs_release_extent_buffer(eb);
4656         return exists;
4657 }
4658
4659 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4660 {
4661         struct extent_buffer *eb =
4662                         container_of(head, struct extent_buffer, rcu_head);
4663
4664         __free_extent_buffer(eb);
4665 }
4666
4667 /* Expects to have eb->eb_lock already held */
4668 static int release_extent_buffer(struct extent_buffer *eb)
4669 {
4670         WARN_ON(atomic_read(&eb->refs) == 0);
4671         if (atomic_dec_and_test(&eb->refs)) {
4672                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
4673                         struct btrfs_fs_info *fs_info = eb->fs_info;
4674
4675                         spin_unlock(&eb->refs_lock);
4676
4677                         spin_lock(&fs_info->buffer_lock);
4678                         radix_tree_delete(&fs_info->buffer_radix,
4679                                           eb->start >> PAGE_CACHE_SHIFT);
4680                         spin_unlock(&fs_info->buffer_lock);
4681                 } else {
4682                         spin_unlock(&eb->refs_lock);
4683                 }
4684
4685                 /* Should be safe to release our pages at this point */
4686                 btrfs_release_extent_buffer_page(eb, 0);
4687                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4688                 return 1;
4689         }
4690         spin_unlock(&eb->refs_lock);
4691
4692         return 0;
4693 }
4694
4695 void free_extent_buffer(struct extent_buffer *eb)
4696 {
4697         int refs;
4698         int old;
4699         if (!eb)
4700                 return;
4701
4702         while (1) {
4703                 refs = atomic_read(&eb->refs);
4704                 if (refs <= 3)
4705                         break;
4706                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
4707                 if (old == refs)
4708                         return;
4709         }
4710
4711         spin_lock(&eb->refs_lock);
4712         if (atomic_read(&eb->refs) == 2 &&
4713             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
4714                 atomic_dec(&eb->refs);
4715
4716         if (atomic_read(&eb->refs) == 2 &&
4717             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4718             !extent_buffer_under_io(eb) &&
4719             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4720                 atomic_dec(&eb->refs);
4721
4722         /*
4723          * I know this is terrible, but it's temporary until we stop tracking
4724          * the uptodate bits and such for the extent buffers.
4725          */
4726         release_extent_buffer(eb);
4727 }
4728
4729 void free_extent_buffer_stale(struct extent_buffer *eb)
4730 {
4731         if (!eb)
4732                 return;
4733
4734         spin_lock(&eb->refs_lock);
4735         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4736
4737         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4738             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4739                 atomic_dec(&eb->refs);
4740         release_extent_buffer(eb);
4741 }
4742
4743 void clear_extent_buffer_dirty(struct extent_buffer *eb)
4744 {
4745         unsigned long i;
4746         unsigned long num_pages;
4747         struct page *page;
4748
4749         num_pages = num_extent_pages(eb->start, eb->len);
4750
4751         for (i = 0; i < num_pages; i++) {
4752                 page = extent_buffer_page(eb, i);
4753                 if (!PageDirty(page))
4754                         continue;
4755
4756                 lock_page(page);
4757                 WARN_ON(!PagePrivate(page));
4758
4759                 clear_page_dirty_for_io(page);
4760                 spin_lock_irq(&page->mapping->tree_lock);
4761                 if (!PageDirty(page)) {
4762                         radix_tree_tag_clear(&page->mapping->page_tree,
4763                                                 page_index(page),
4764                                                 PAGECACHE_TAG_DIRTY);
4765                 }
4766                 spin_unlock_irq(&page->mapping->tree_lock);
4767                 ClearPageError(page);
4768                 unlock_page(page);
4769         }
4770         WARN_ON(atomic_read(&eb->refs) == 0);
4771 }
4772
4773 int set_extent_buffer_dirty(struct extent_buffer *eb)
4774 {
4775         unsigned long i;
4776         unsigned long num_pages;
4777         int was_dirty = 0;
4778
4779         check_buffer_tree_ref(eb);
4780
4781         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4782
4783         num_pages = num_extent_pages(eb->start, eb->len);
4784         WARN_ON(atomic_read(&eb->refs) == 0);
4785         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4786
4787         for (i = 0; i < num_pages; i++)
4788                 set_page_dirty(extent_buffer_page(eb, i));
4789         return was_dirty;
4790 }
4791
4792 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4793 {
4794         unsigned long i;
4795         struct page *page;
4796         unsigned long num_pages;
4797
4798         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4799         num_pages = num_extent_pages(eb->start, eb->len);
4800         for (i = 0; i < num_pages; i++) {
4801                 page = extent_buffer_page(eb, i);
4802                 if (page)
4803                         ClearPageUptodate(page);
4804         }
4805         return 0;
4806 }
4807
4808 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4809 {
4810         unsigned long i;
4811         struct page *page;
4812         unsigned long num_pages;
4813
4814         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4815         num_pages = num_extent_pages(eb->start, eb->len);
4816         for (i = 0; i < num_pages; i++) {
4817                 page = extent_buffer_page(eb, i);
4818                 SetPageUptodate(page);
4819         }
4820         return 0;
4821 }
4822
4823 int extent_buffer_uptodate(struct extent_buffer *eb)
4824 {
4825         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4826 }
4827
4828 int read_extent_buffer_pages(struct extent_io_tree *tree,
4829                              struct extent_buffer *eb, u64 start, int wait,
4830                              get_extent_t *get_extent, int mirror_num)
4831 {
4832         unsigned long i;
4833         unsigned long start_i;
4834         struct page *page;
4835         int err;
4836         int ret = 0;
4837         int locked_pages = 0;
4838         int all_uptodate = 1;
4839         unsigned long num_pages;
4840         unsigned long num_reads = 0;
4841         struct bio *bio = NULL;
4842         unsigned long bio_flags = 0;
4843
4844         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4845                 return 0;
4846
4847         if (start) {
4848                 WARN_ON(start < eb->start);
4849                 start_i = (start >> PAGE_CACHE_SHIFT) -
4850                         (eb->start >> PAGE_CACHE_SHIFT);
4851         } else {
4852                 start_i = 0;
4853         }
4854
4855         num_pages = num_extent_pages(eb->start, eb->len);
4856         for (i = start_i; i < num_pages; i++) {
4857                 page = extent_buffer_page(eb, i);
4858                 if (wait == WAIT_NONE) {
4859                         if (!trylock_page(page))
4860                                 goto unlock_exit;
4861                 } else {
4862                         lock_page(page);
4863                 }
4864                 locked_pages++;
4865                 if (!PageUptodate(page)) {
4866                         num_reads++;
4867                         all_uptodate = 0;
4868                 }
4869         }
4870         if (all_uptodate) {
4871                 if (start_i == 0)
4872                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4873                 goto unlock_exit;
4874         }
4875
4876         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4877         eb->read_mirror = 0;
4878         atomic_set(&eb->io_pages, num_reads);
4879         for (i = start_i; i < num_pages; i++) {
4880                 page = extent_buffer_page(eb, i);
4881                 if (!PageUptodate(page)) {
4882                         ClearPageError(page);
4883                         err = __extent_read_full_page(tree, page,
4884                                                       get_extent, &bio,
4885                                                       mirror_num, &bio_flags,
4886                                                       READ | REQ_META);
4887                         if (err)
4888                                 ret = err;
4889                 } else {
4890                         unlock_page(page);
4891                 }
4892         }
4893
4894         if (bio) {
4895                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
4896                                      bio_flags);
4897                 if (err)
4898                         return err;
4899         }
4900
4901         if (ret || wait != WAIT_COMPLETE)
4902                 return ret;
4903
4904         for (i = start_i; i < num_pages; i++) {
4905                 page = extent_buffer_page(eb, i);
4906                 wait_on_page_locked(page);
4907                 if (!PageUptodate(page))
4908                         ret = -EIO;
4909         }
4910
4911         return ret;
4912
4913 unlock_exit:
4914         i = start_i;
4915         while (locked_pages > 0) {
4916                 page = extent_buffer_page(eb, i);
4917                 i++;
4918                 unlock_page(page);
4919                 locked_pages--;
4920         }
4921         return ret;
4922 }
4923
4924 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4925                         unsigned long start,
4926                         unsigned long len)
4927 {
4928         size_t cur;
4929         size_t offset;
4930         struct page *page;
4931         char *kaddr;
4932         char *dst = (char *)dstv;
4933         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4934         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4935
4936         WARN_ON(start > eb->len);
4937         WARN_ON(start + len > eb->start + eb->len);
4938
4939         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
4940
4941         while (len > 0) {
4942                 page = extent_buffer_page(eb, i);
4943
4944                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4945                 kaddr = page_address(page);
4946                 memcpy(dst, kaddr + offset, cur);
4947
4948                 dst += cur;
4949                 len -= cur;
4950                 offset = 0;
4951                 i++;
4952         }
4953 }
4954
4955 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4956                                unsigned long min_len, char **map,
4957                                unsigned long *map_start,
4958                                unsigned long *map_len)
4959 {
4960         size_t offset = start & (PAGE_CACHE_SIZE - 1);
4961         char *kaddr;
4962         struct page *p;
4963         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4964         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4965         unsigned long end_i = (start_offset + start + min_len - 1) >>
4966                 PAGE_CACHE_SHIFT;
4967
4968         if (i != end_i)
4969                 return -EINVAL;
4970
4971         if (i == 0) {
4972                 offset = start_offset;
4973                 *map_start = 0;
4974         } else {
4975                 offset = 0;
4976                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4977         }
4978
4979         if (start + min_len > eb->len) {
4980                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4981                        "wanted %lu %lu\n",
4982                        eb->start, eb->len, start, min_len);
4983                 return -EINVAL;
4984         }
4985
4986         p = extent_buffer_page(eb, i);
4987         kaddr = page_address(p);
4988         *map = kaddr + offset;
4989         *map_len = PAGE_CACHE_SIZE - offset;
4990         return 0;
4991 }
4992
4993 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4994                           unsigned long start,
4995                           unsigned long len)
4996 {
4997         size_t cur;
4998         size_t offset;
4999         struct page *page;
5000         char *kaddr;
5001         char *ptr = (char *)ptrv;
5002         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5003         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5004         int ret = 0;
5005
5006         WARN_ON(start > eb->len);
5007         WARN_ON(start + len > eb->start + eb->len);
5008
5009         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5010
5011         while (len > 0) {
5012                 page = extent_buffer_page(eb, i);
5013
5014                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5015
5016                 kaddr = page_address(page);
5017                 ret = memcmp(ptr, kaddr + offset, cur);
5018                 if (ret)
5019                         break;
5020
5021                 ptr += cur;
5022                 len -= cur;
5023                 offset = 0;
5024                 i++;
5025         }
5026         return ret;
5027 }
5028
5029 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5030                          unsigned long start, unsigned long len)
5031 {
5032         size_t cur;
5033         size_t offset;
5034         struct page *page;
5035         char *kaddr;
5036         char *src = (char *)srcv;
5037         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5038         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5039
5040         WARN_ON(start > eb->len);
5041         WARN_ON(start + len > eb->start + eb->len);
5042
5043         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5044
5045         while (len > 0) {
5046                 page = extent_buffer_page(eb, i);
5047                 WARN_ON(!PageUptodate(page));
5048
5049                 cur = min(len, PAGE_CACHE_SIZE - offset);
5050                 kaddr = page_address(page);
5051                 memcpy(kaddr + offset, src, cur);
5052
5053                 src += cur;
5054                 len -= cur;
5055                 offset = 0;
5056                 i++;
5057         }
5058 }
5059
5060 void memset_extent_buffer(struct extent_buffer *eb, char c,
5061                           unsigned long start, unsigned long len)
5062 {
5063         size_t cur;
5064         size_t offset;
5065         struct page *page;
5066         char *kaddr;
5067         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5068         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5069
5070         WARN_ON(start > eb->len);
5071         WARN_ON(start + len > eb->start + eb->len);
5072
5073         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5074
5075         while (len > 0) {
5076                 page = extent_buffer_page(eb, i);
5077                 WARN_ON(!PageUptodate(page));
5078
5079                 cur = min(len, PAGE_CACHE_SIZE - offset);
5080                 kaddr = page_address(page);
5081                 memset(kaddr + offset, c, cur);
5082
5083                 len -= cur;
5084                 offset = 0;
5085                 i++;
5086         }
5087 }
5088
5089 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5090                         unsigned long dst_offset, unsigned long src_offset,
5091                         unsigned long len)
5092 {
5093         u64 dst_len = dst->len;
5094         size_t cur;
5095         size_t offset;
5096         struct page *page;
5097         char *kaddr;
5098         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5099         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5100
5101         WARN_ON(src->len != dst_len);
5102
5103         offset = (start_offset + dst_offset) &
5104                 (PAGE_CACHE_SIZE - 1);
5105
5106         while (len > 0) {
5107                 page = extent_buffer_page(dst, i);
5108                 WARN_ON(!PageUptodate(page));
5109
5110                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5111
5112                 kaddr = page_address(page);
5113                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5114
5115                 src_offset += cur;
5116                 len -= cur;
5117                 offset = 0;
5118                 i++;
5119         }
5120 }
5121
5122 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5123 {
5124         unsigned long distance = (src > dst) ? src - dst : dst - src;
5125         return distance < len;
5126 }
5127
5128 static void copy_pages(struct page *dst_page, struct page *src_page,
5129                        unsigned long dst_off, unsigned long src_off,
5130                        unsigned long len)
5131 {
5132         char *dst_kaddr = page_address(dst_page);
5133         char *src_kaddr;
5134         int must_memmove = 0;
5135
5136         if (dst_page != src_page) {
5137                 src_kaddr = page_address(src_page);
5138         } else {
5139                 src_kaddr = dst_kaddr;
5140                 if (areas_overlap(src_off, dst_off, len))
5141                         must_memmove = 1;
5142         }
5143
5144         if (must_memmove)
5145                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5146         else
5147                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5148 }
5149
5150 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5151                            unsigned long src_offset, unsigned long len)
5152 {
5153         size_t cur;
5154         size_t dst_off_in_page;
5155         size_t src_off_in_page;
5156         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5157         unsigned long dst_i;
5158         unsigned long src_i;
5159
5160         if (src_offset + len > dst->len) {
5161                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5162                        "len %lu dst len %lu\n", src_offset, len, dst->len);
5163                 BUG_ON(1);
5164         }
5165         if (dst_offset + len > dst->len) {
5166                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5167                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
5168                 BUG_ON(1);
5169         }
5170
5171         while (len > 0) {
5172                 dst_off_in_page = (start_offset + dst_offset) &
5173                         (PAGE_CACHE_SIZE - 1);
5174                 src_off_in_page = (start_offset + src_offset) &
5175                         (PAGE_CACHE_SIZE - 1);
5176
5177                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5178                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5179
5180                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5181                                                src_off_in_page));
5182                 cur = min_t(unsigned long, cur,
5183                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5184
5185                 copy_pages(extent_buffer_page(dst, dst_i),
5186                            extent_buffer_page(dst, src_i),
5187                            dst_off_in_page, src_off_in_page, cur);
5188
5189                 src_offset += cur;
5190                 dst_offset += cur;
5191                 len -= cur;
5192         }
5193 }
5194
5195 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5196                            unsigned long src_offset, unsigned long len)
5197 {
5198         size_t cur;
5199         size_t dst_off_in_page;
5200         size_t src_off_in_page;
5201         unsigned long dst_end = dst_offset + len - 1;
5202         unsigned long src_end = src_offset + len - 1;
5203         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5204         unsigned long dst_i;
5205         unsigned long src_i;
5206
5207         if (src_offset + len > dst->len) {
5208                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5209                        "len %lu len %lu\n", src_offset, len, dst->len);
5210                 BUG_ON(1);
5211         }
5212         if (dst_offset + len > dst->len) {
5213                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5214                        "len %lu len %lu\n", dst_offset, len, dst->len);
5215                 BUG_ON(1);
5216         }
5217         if (dst_offset < src_offset) {
5218                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5219                 return;
5220         }
5221         while (len > 0) {
5222                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5223                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5224
5225                 dst_off_in_page = (start_offset + dst_end) &
5226                         (PAGE_CACHE_SIZE - 1);
5227                 src_off_in_page = (start_offset + src_end) &
5228                         (PAGE_CACHE_SIZE - 1);
5229
5230                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5231                 cur = min(cur, dst_off_in_page + 1);
5232                 copy_pages(extent_buffer_page(dst, dst_i),
5233                            extent_buffer_page(dst, src_i),
5234                            dst_off_in_page - cur + 1,
5235                            src_off_in_page - cur + 1, cur);
5236
5237                 dst_end -= cur;
5238                 src_end -= cur;
5239                 len -= cur;
5240         }
5241 }
5242
5243 int try_release_extent_buffer(struct page *page)
5244 {
5245         struct extent_buffer *eb;
5246
5247         /*
5248          * We need to make sure noboody is attaching this page to an eb right
5249          * now.
5250          */
5251         spin_lock(&page->mapping->private_lock);
5252         if (!PagePrivate(page)) {
5253                 spin_unlock(&page->mapping->private_lock);
5254                 return 1;
5255         }
5256
5257         eb = (struct extent_buffer *)page->private;
5258         BUG_ON(!eb);
5259
5260         /*
5261          * This is a little awful but should be ok, we need to make sure that
5262          * the eb doesn't disappear out from under us while we're looking at
5263          * this page.
5264          */
5265         spin_lock(&eb->refs_lock);
5266         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5267                 spin_unlock(&eb->refs_lock);
5268                 spin_unlock(&page->mapping->private_lock);
5269                 return 0;
5270         }
5271         spin_unlock(&page->mapping->private_lock);
5272
5273         /*
5274          * If tree ref isn't set then we know the ref on this eb is a real ref,
5275          * so just return, this page will likely be freed soon anyway.
5276          */
5277         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5278                 spin_unlock(&eb->refs_lock);
5279                 return 0;
5280         }
5281
5282         return release_extent_buffer(eb);
5283 }