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,
 233                                    struct rb_node *search_start,
 234                                    u64 offset,
 235                                    struct rb_node *node,
 236                                    struct rb_node ***p_in,
 237                                    struct rb_node **parent_in)
 238 {
 239         struct rb_node **p;
 240         struct rb_node *parent = NULL;
 241         struct tree_entry *entry;
 242
 243         if (p_in && parent_in) {
 244                 p = *p_in;
 245                 parent = *parent_in;
 246                 goto do_insert;
 247         }
 248
 249         p = search_start ? &search_start : &root->rb_node;
 250         while (*p) {
 251                 parent = *p;
 252                 entry = rb_entry(parent, struct tree_entry, rb_node);
 253
 254                 if (offset < entry->start)
 255                         p = &(*p)->rb_left;
 256                 else if (offset > entry->end)
 257                         p = &(*p)->rb_right;
 258                 else
 259                         return parent;
 260         }
 261
 262 do_insert:
 263         rb_link_node(node, parent, p);
 264         rb_insert_color(node, root);
 265         return NULL;
 266 }
 267
 268 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 269                                       struct rb_node **prev_ret,
 270                                       struct rb_node **next_ret,
 271                                       struct rb_node ***p_ret,
 272                                       struct rb_node **parent_ret)
 273 {
 274         struct rb_root *root = &tree->state;
 275         struct rb_node **n = &root->rb_node;
 276         struct rb_node *prev = NULL;
 277         struct rb_node *orig_prev = NULL;
 278         struct tree_entry *entry;
 279         struct tree_entry *prev_entry = NULL;
 280
 281         while (*n) {
 282                 prev = *n;
 283                 entry = rb_entry(prev, struct tree_entry, rb_node);
 284                 prev_entry = entry;
 285
 286                 if (offset < entry->start)
 287                         n = &(*n)->rb_left;
 288                 else if (offset > entry->end)
 289                         n = &(*n)->rb_right;
 290                 else
 291                         return *n;
 292         }
 293
 294         if (p_ret)
 295                 *p_ret = n;
 296         if (parent_ret)
 297                 *parent_ret = prev;
 298
 299         if (prev_ret) {
 300                 orig_prev = prev;
 301                 while (prev && offset > prev_entry->end) {
 302                         prev = rb_next(prev);
 303                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 304                 }
 305                 *prev_ret = prev;
 306                 prev = orig_prev;
 307         }
 308
 309         if (next_ret) {
 310                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 311                 while (prev && offset < prev_entry->start) {
 312                         prev = rb_prev(prev);
 313                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 314                 }
 315                 *next_ret = prev;
 316         }
 317         return NULL;
 318 }
 319
 320 static inline struct rb_node *
 321 tree_search_for_insert(struct extent_io_tree *tree,
 322                        u64 offset,
 323                        struct rb_node ***p_ret,
 324                        struct rb_node **parent_ret)
 325 {
 326         struct rb_node *prev = NULL;
 327         struct rb_node *ret;
 328
 329         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 330         if (!ret)
 331                 return prev;
 332         return ret;
 333 }
 334
 335 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 336                                           u64 offset)
 337 {
 338         return tree_search_for_insert(tree, offset, NULL, NULL);
 339 }
 340
 341 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 342                      struct extent_state *other)
 343 {
 344         if (tree->ops && tree->ops->merge_extent_hook)
 345                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 346                                              other);
 347 }
 348
 349 /*
 350  * utility function to look for merge candidates inside a given range.
 351  * Any extents with matching state are merged together into a single
 352  * extent in the tree.  Extents with EXTENT_IO in their state field
 353  * are not merged because the end_io handlers need to be able to do
 354  * operations on them without sleeping (or doing allocations/splits).
 355  *
 356  * This should be called with the tree lock held.
 357  */
 358 static void merge_state(struct extent_io_tree *tree,
 359                         struct extent_state *state)
 360 {
 361         struct extent_state *other;
 362         struct rb_node *other_node;
 363
 364         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 365                 return;
 366
 367         other_node = rb_prev(&state->rb_node);
 368         if (other_node) {
 369                 other = rb_entry(other_node, struct extent_state, rb_node);
 370                 if (other->end == state->start - 1 &&
 371                     other->state == state->state) {
 372                         merge_cb(tree, state, other);
 373                         state->start = other->start;
 374                         other->tree = NULL;
 375                         rb_erase(&other->rb_node, &tree->state);
 376                         free_extent_state(other);
 377                 }
 378         }
 379         other_node = rb_next(&state->rb_node);
 380         if (other_node) {
 381                 other = rb_entry(other_node, struct extent_state, rb_node);
 382                 if (other->start == state->end + 1 &&
 383                     other->state == state->state) {
 384                         merge_cb(tree, state, other);
 385                         state->end = other->end;
 386                         other->tree = NULL;
 387                         rb_erase(&other->rb_node, &tree->state);
 388                         free_extent_state(other);
 389                 }
 390         }
 391 }
 392
 393 static void set_state_cb(struct extent_io_tree *tree,
 394                          struct extent_state *state, unsigned long *bits)
 395 {
 396         if (tree->ops && tree->ops->set_bit_hook)
 397                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 398 }
 399
 400 static void clear_state_cb(struct extent_io_tree *tree,
 401                            struct extent_state *state, unsigned long *bits)
 402 {
 403         if (tree->ops && tree->ops->clear_bit_hook)
 404                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 405 }
 406
 407 static void set_state_bits(struct extent_io_tree *tree,
 408                            struct extent_state *state, unsigned long *bits);
 409
 410 /*
 411  * insert an extent_state struct into the tree.  'bits' are set on the
 412  * struct before it is inserted.
 413  *
 414  * This may return -EEXIST if the extent is already there, in which case the
 415  * state struct is freed.
 416  *
 417  * The tree lock is not taken internally.  This is a utility function and
 418  * probably isn't what you want to call (see set/clear_extent_bit).
 419  */
 420 static int insert_state(struct extent_io_tree *tree,
 421                         struct extent_state *state, u64 start, u64 end,
 422                         struct rb_node ***p,
 423                         struct rb_node **parent,
 424                         unsigned long *bits)
 425 {
 426         struct rb_node *node;
 427
 428         if (end < start)
 429                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 430                        end, start);
 431         state->start = start;
 432         state->end = end;
 433
 434         set_state_bits(tree, state, bits);
 435
 436         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 437         if (node) {
 438                 struct extent_state *found;
 439                 found = rb_entry(node, struct extent_state, rb_node);
 440                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
 441                        "%llu %llu\n",
 442                        found->start, found->end, start, end);
 443                 return -EEXIST;
 444         }
 445         state->tree = tree;
 446         merge_state(tree, state);
 447         return 0;
 448 }
 449
 450 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 451                      u64 split)
 452 {
 453         if (tree->ops && tree->ops->split_extent_hook)
 454                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 455 }
 456
 457 /*
 458  * split a given extent state struct in two, inserting the preallocated
 459  * struct 'prealloc' as the newly created second half.  'split' indicates an
 460  * offset inside 'orig' where it should be split.
 461  *
 462  * Before calling,
 463  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 464  * are two extent state structs in the tree:
 465  * prealloc: [orig->start, split - 1]
 466  * orig: [ split, orig->end ]
 467  *
 468  * The tree locks are not taken by this function. They need to be held
 469  * by the caller.
 470  */
 471 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 472                        struct extent_state *prealloc, u64 split)
 473 {
 474         struct rb_node *node;
 475
 476         split_cb(tree, orig, split);
 477
 478         prealloc->start = orig->start;
 479         prealloc->end = split - 1;
 480         prealloc->state = orig->state;
 481         orig->start = split;
 482
 483         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 484                            &prealloc->rb_node, NULL, NULL);
 485         if (node) {
 486                 free_extent_state(prealloc);
 487                 return -EEXIST;
 488         }
 489         prealloc->tree = tree;
 490         return 0;
 491 }
 492
 493 static struct extent_state *next_state(struct extent_state *state)
 494 {
 495         struct rb_node *next = rb_next(&state->rb_node);
 496         if (next)
 497                 return rb_entry(next, struct extent_state, rb_node);
 498         else
 499                 return NULL;
 500 }
 501
 502 /*
 503  * utility function to clear some bits in an extent state struct.
 504  * it will optionally wake up any one waiting on this state (wake == 1).
 505  *
 506  * If no bits are set on the state struct after clearing things, the
 507  * struct is freed and removed from the tree
 508  */
 509 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 510                                             struct extent_state *state,
 511                                             unsigned long *bits, int wake)
 512 {
 513         struct extent_state *next;
 514         unsigned long bits_to_clear = *bits & ~EXTENT_CTLBITS;
 515
 516         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 517                 u64 range = state->end - state->start + 1;
 518                 WARN_ON(range > tree->dirty_bytes);
 519                 tree->dirty_bytes -= range;
 520         }
 521         clear_state_cb(tree, state, bits);
 522         state->state &= ~bits_to_clear;
 523         if (wake)
 524                 wake_up(&state->wq);
 525         if (state->state == 0) {
 526                 next = next_state(state);
 527                 if (state->tree) {
 528                         rb_erase(&state->rb_node, &tree->state);
 529                         state->tree = NULL;
 530                         free_extent_state(state);
 531                 } else {
 532                         WARN_ON(1);
 533                 }
 534         } else {
 535                 merge_state(tree, state);
 536                 next = next_state(state);
 537         }
 538         return next;
 539 }
 540
 541 static struct extent_state *
 542 alloc_extent_state_atomic(struct extent_state *prealloc)
 543 {
 544         if (!prealloc)
 545                 prealloc = alloc_extent_state(GFP_ATOMIC);
 546
 547         return prealloc;
 548 }
 549
 550 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 551 {
 552         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 553                     "Extent tree was modified by another "
 554                     "thread while locked.");
 555 }
 556
 557 /*
 558  * clear some bits on a range in the tree.  This may require splitting
 559  * or inserting elements in the tree, so the gfp mask is used to
 560  * indicate which allocations or sleeping are allowed.
 561  *
 562  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 563  * the given range from the tree regardless of state (ie for truncate).
 564  *
 565  * the range [start, end] is inclusive.
 566  *
 567  * This takes the tree lock, and returns 0 on success and < 0 on error.
 568  */
 569 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 570                      unsigned long bits, int wake, int delete,
 571                      struct extent_state **cached_state,
 572                      gfp_t mask)
 573 {
 574         struct extent_state *state;
 575         struct extent_state *cached;
 576         struct extent_state *prealloc = NULL;
 577         struct rb_node *node;
 578         u64 last_end;
 579         int err;
 580         int clear = 0;
 581
 582         btrfs_debug_check_extent_io_range(tree, start, end);
 583
 584         if (bits & EXTENT_DELALLOC)
 585                 bits |= EXTENT_NORESERVE;
 586
 587         if (delete)
 588                 bits |= ~EXTENT_CTLBITS;
 589         bits |= EXTENT_FIRST_DELALLOC;
 590
 591         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 592                 clear = 1;
 593 again:
 594         if (!prealloc && (mask & __GFP_WAIT)) {
 595                 prealloc = alloc_extent_state(mask);
 596                 if (!prealloc)
 597                         return -ENOMEM;
 598         }
 599
 600         spin_lock(&tree->lock);
 601         if (cached_state) {
 602                 cached = *cached_state;
 603
 604                 if (clear) {
 605                         *cached_state = NULL;
 606                         cached_state = NULL;
 607                 }
 608
 609                 if (cached && cached->tree && cached->start <= start &&
 610                     cached->end > start) {
 611                         if (clear)
 612                                 atomic_dec(&cached->refs);
 613                         state = cached;
 614                         goto hit_next;
 615                 }
 616                 if (clear)
 617                         free_extent_state(cached);
 618         }
 619         /*
 620          * this search will find the extents that end after
 621          * our range starts
 622          */
 623         node = tree_search(tree, start);
 624         if (!node)
 625                 goto out;
 626         state = rb_entry(node, struct extent_state, rb_node);
 627 hit_next:
 628         if (state->start > end)
 629                 goto out;
 630         WARN_ON(state->end < start);
 631         last_end = state->end;
 632
 633         /* the state doesn't have the wanted bits, go ahead */
 634         if (!(state->state & bits)) {
 635                 state = next_state(state);
 636                 goto next;
 637         }
 638
 639         /*
 640          *     | ---- desired range ---- |
 641          *  | state | or
 642          *  | ------------- state -------------- |
 643          *
 644          * We need to split the extent we found, and may flip
 645          * bits on second half.
 646          *
 647          * If the extent we found extends past our range, we
 648          * just split and search again.  It'll get split again
 649          * the next time though.
 650          *
 651          * If the extent we found is inside our range, we clear
 652          * the desired bit on it.
 653          */
 654
 655         if (state->start < start) {
 656                 prealloc = alloc_extent_state_atomic(prealloc);
 657                 BUG_ON(!prealloc);
 658                 err = split_state(tree, state, prealloc, start);
 659                 if (err)
 660                         extent_io_tree_panic(tree, err);
 661
 662                 prealloc = NULL;
 663                 if (err)
 664                         goto out;
 665                 if (state->end <= end) {
 666                         state = clear_state_bit(tree, state, &bits, wake);
 667                         goto next;
 668                 }
 669                 goto search_again;
 670         }
 671         /*
 672          * | ---- desired range ---- |
 673          *                        | state |
 674          * We need to split the extent, and clear the bit
 675          * on the first half
 676          */
 677         if (state->start <= end && state->end > end) {
 678                 prealloc = alloc_extent_state_atomic(prealloc);
 679                 BUG_ON(!prealloc);
 680                 err = split_state(tree, state, prealloc, end + 1);
 681                 if (err)
 682                         extent_io_tree_panic(tree, err);
 683
 684                 if (wake)
 685                         wake_up(&state->wq);
 686
 687                 clear_state_bit(tree, prealloc, &bits, wake);
 688
 689                 prealloc = NULL;
 690                 goto out;
 691         }
 692
 693         state = clear_state_bit(tree, state, &bits, wake);
 694 next:
 695         if (last_end == (u64)-1)
 696                 goto out;
 697         start = last_end + 1;
 698         if (start <= end && state && !need_resched())
 699                 goto hit_next;
 700         goto search_again;
 701
 702 out:
 703         spin_unlock(&tree->lock);
 704         if (prealloc)
 705                 free_extent_state(prealloc);
 706
 707         return 0;
 708
 709 search_again:
 710         if (start > end)
 711                 goto out;
 712         spin_unlock(&tree->lock);
 713         if (mask & __GFP_WAIT)
 714                 cond_resched();
 715         goto again;
 716 }
 717
 718 static void wait_on_state(struct extent_io_tree *tree,
 719                           struct extent_state *state)
 720                 __releases(tree->lock)
 721                 __acquires(tree->lock)
 722 {
 723         DEFINE_WAIT(wait);
 724         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 725         spin_unlock(&tree->lock);
 726         schedule();
 727         spin_lock(&tree->lock);
 728         finish_wait(&state->wq, &wait);
 729 }
 730
 731 /*
 732  * waits for one or more bits to clear on a range in the state tree.
 733  * The range [start, end] is inclusive.
 734  * The tree lock is taken by this function
 735  */
 736 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 737                             unsigned long bits)
 738 {
 739         struct extent_state *state;
 740         struct rb_node *node;
 741
 742         btrfs_debug_check_extent_io_range(tree, start, end);
 743
 744         spin_lock(&tree->lock);
 745 again:
 746         while (1) {
 747                 /*
 748                  * this search will find all the extents that end after
 749                  * our range starts
 750                  */
 751                 node = tree_search(tree, start);
 752 process_node:
 753                 if (!node)
 754                         break;
 755
 756                 state = rb_entry(node, struct extent_state, rb_node);
 757
 758                 if (state->start > end)
 759                         goto out;
 760
 761                 if (state->state & bits) {
 762                         start = state->start;
 763                         atomic_inc(&state->refs);
 764                         wait_on_state(tree, state);
 765                         free_extent_state(state);
 766                         goto again;
 767                 }
 768                 start = state->end + 1;
 769
 770                 if (start > end)
 771                         break;
 772
 773                 if (!cond_resched_lock(&tree->lock)) {
 774                         node = rb_next(node);
 775                         goto process_node;
 776                 }
 777         }
 778 out:
 779         spin_unlock(&tree->lock);
 780 }
 781
 782 static void set_state_bits(struct extent_io_tree *tree,
 783                            struct extent_state *state,
 784                            unsigned long *bits)
 785 {
 786         unsigned long bits_to_set = *bits & ~EXTENT_CTLBITS;
 787
 788         set_state_cb(tree, state, bits);
 789         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 790                 u64 range = state->end - state->start + 1;
 791                 tree->dirty_bytes += range;
 792         }
 793         state->state |= bits_to_set;
 794 }
 795
 796 static void cache_state(struct extent_state *state,
 797                         struct extent_state **cached_ptr)
 798 {
 799         if (cached_ptr && !(*cached_ptr)) {
 800                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
 801                         *cached_ptr = state;
 802                         atomic_inc(&state->refs);
 803                 }
 804         }
 805 }
 806
 807 /*
 808  * set some bits on a range in the tree.  This may require allocations or
 809  * sleeping, so the gfp mask is used to indicate what is allowed.
 810  *
 811  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 812  * part of the range already has the desired bits set.  The start of the
 813  * existing range is returned in failed_start in this case.
 814  *
 815  * [start, end] is inclusive This takes the tree lock.
 816  */
 817
 818 static int __must_check
 819 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 820                  unsigned long bits, unsigned long exclusive_bits,
 821                  u64 *failed_start, struct extent_state **cached_state,
 822                  gfp_t mask)
 823 {
 824         struct extent_state *state;
 825         struct extent_state *prealloc = NULL;
 826         struct rb_node *node;
 827         struct rb_node **p;
 828         struct rb_node *parent;
 829         int err = 0;
 830         u64 last_start;
 831         u64 last_end;
 832
 833         btrfs_debug_check_extent_io_range(tree, start, end);
 834
 835         bits |= EXTENT_FIRST_DELALLOC;
 836 again:
 837         if (!prealloc && (mask & __GFP_WAIT)) {
 838                 prealloc = alloc_extent_state(mask);
 839                 BUG_ON(!prealloc);
 840         }
 841
 842         spin_lock(&tree->lock);
 843         if (cached_state && *cached_state) {
 844                 state = *cached_state;
 845                 if (state->start <= start && state->end > start &&
 846                     state->tree) {
 847                         node = &state->rb_node;
 848                         goto hit_next;
 849                 }
 850         }
 851         /*
 852          * this search will find all the extents that end after
 853          * our range starts.
 854          */
 855         node = tree_search_for_insert(tree, start, &p, &parent);
 856         if (!node) {
 857                 prealloc = alloc_extent_state_atomic(prealloc);
 858                 BUG_ON(!prealloc);
 859                 err = insert_state(tree, prealloc, start, end,
 860                                    &p, &parent, &bits);
 861                 if (err)
 862                         extent_io_tree_panic(tree, err);
 863
 864                 cache_state(prealloc, cached_state);
 865                 prealloc = NULL;
 866                 goto out;
 867         }
 868         state = rb_entry(node, struct extent_state, rb_node);
 869 hit_next:
 870         last_start = state->start;
 871         last_end = state->end;
 872
 873         /*
 874          * | ---- desired range ---- |
 875          * | state |
 876          *
 877          * Just lock what we found and keep going
 878          */
 879         if (state->start == start && state->end <= end) {
 880                 if (state->state & exclusive_bits) {
 881                         *failed_start = state->start;
 882                         err = -EEXIST;
 883                         goto out;
 884                 }
 885
 886                 set_state_bits(tree, state, &bits);
 887                 cache_state(state, cached_state);
 888                 merge_state(tree, state);
 889                 if (last_end == (u64)-1)
 890                         goto out;
 891                 start = last_end + 1;
 892                 state = next_state(state);
 893                 if (start < end && state && state->start == start &&
 894                     !need_resched())
 895                         goto hit_next;
 896                 goto search_again;
 897         }
 898
 899         /*
 900          *     | ---- desired range ---- |
 901          * | state |
 902          *   or
 903          * | ------------- state -------------- |
 904          *
 905          * We need to split the extent we found, and may flip bits on
 906          * second half.
 907          *
 908          * If the extent we found extends past our
 909          * range, we just split and search again.  It'll get split
 910          * again the next time though.
 911          *
 912          * If the extent we found is inside our range, we set the
 913          * desired bit on it.
 914          */
 915         if (state->start < start) {
 916                 if (state->state & exclusive_bits) {
 917                         *failed_start = start;
 918                         err = -EEXIST;
 919                         goto out;
 920                 }
 921
 922                 prealloc = alloc_extent_state_atomic(prealloc);
 923                 BUG_ON(!prealloc);
 924                 err = split_state(tree, state, prealloc, start);
 925                 if (err)
 926                         extent_io_tree_panic(tree, err);
 927
 928                 prealloc = NULL;
 929                 if (err)
 930                         goto out;
 931                 if (state->end <= end) {
 932                         set_state_bits(tree, state, &bits);
 933                         cache_state(state, cached_state);
 934                         merge_state(tree, state);
 935                         if (last_end == (u64)-1)
 936                                 goto out;
 937                         start = last_end + 1;
 938                         state = next_state(state);
 939                         if (start < end && state && state->start == start &&
 940                             !need_resched())
 941                                 goto hit_next;
 942                 }
 943                 goto search_again;
 944         }
 945         /*
 946          * | ---- desired range ---- |
 947          *     | state | or               | state |
 948          *
 949          * There's a hole, we need to insert something in it and
 950          * ignore the extent we found.
 951          */
 952         if (state->start > start) {
 953                 u64 this_end;
 954                 if (end < last_start)
 955                         this_end = end;
 956                 else
 957                         this_end = last_start - 1;
 958
 959                 prealloc = alloc_extent_state_atomic(prealloc);
 960                 BUG_ON(!prealloc);
 961
 962                 /*
 963                  * Avoid to free 'prealloc' if it can be merged with
 964                  * the later extent.
 965                  */
 966                 err = insert_state(tree, prealloc, start, this_end,
 967                                    NULL, NULL, &bits);
 968                 if (err)
 969                         extent_io_tree_panic(tree, err);
 970
 971                 cache_state(prealloc, cached_state);
 972                 prealloc = NULL;
 973                 start = this_end + 1;
 974                 goto search_again;
 975         }
 976         /*
 977          * | ---- desired range ---- |
 978          *                        | state |
 979          * We need to split the extent, and set the bit
 980          * on the first half
 981          */
 982         if (state->start <= end && state->end > end) {
 983                 if (state->state & exclusive_bits) {
 984                         *failed_start = start;
 985                         err = -EEXIST;
 986                         goto out;
 987                 }
 988
 989                 prealloc = alloc_extent_state_atomic(prealloc);
 990                 BUG_ON(!prealloc);
 991                 err = split_state(tree, state, prealloc, end + 1);
 992                 if (err)
 993                         extent_io_tree_panic(tree, err);
 994
 995                 set_state_bits(tree, prealloc, &bits);
 996                 cache_state(prealloc, cached_state);
 997                 merge_state(tree, prealloc);
 998                 prealloc = NULL;
 999                 goto out;
1000         }
1001
1002         goto search_again;
1003
1004 out:
1005         spin_unlock(&tree->lock);
1006         if (prealloc)
1007                 free_extent_state(prealloc);
1008
1009         return err;
1010
1011 search_again:
1012         if (start > end)
1013                 goto out;
1014         spin_unlock(&tree->lock);
1015         if (mask & __GFP_WAIT)
1016                 cond_resched();
1017         goto again;
1018 }
1019
1020 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1021                    unsigned long bits, u64 * failed_start,
1022                    struct extent_state **cached_state, gfp_t mask)
1023 {
1024         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1025                                 cached_state, mask);
1026 }
1027
1028
1029 /**
1030  * convert_extent_bit - convert all bits in a given range from one bit to
1031  *                      another
1032  * @tree:       the io tree to search
1033  * @start:      the start offset in bytes
1034  * @end:        the end offset in bytes (inclusive)
1035  * @bits:       the bits to set in this range
1036  * @clear_bits: the bits to clear in this range
1037  * @cached_state:       state that we're going to cache
1038  * @mask:       the allocation mask
1039  *
1040  * This will go through and set bits for the given range.  If any states exist
1041  * already in this range they are set with the given bit and cleared of the
1042  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1043  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1044  * boundary bits like LOCK.
1045  */
1046 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1047                        unsigned long bits, unsigned long clear_bits,
1048                        struct extent_state **cached_state, gfp_t mask)
1049 {
1050         struct extent_state *state;
1051         struct extent_state *prealloc = NULL;
1052         struct rb_node *node;
1053         struct rb_node **p;
1054         struct rb_node *parent;
1055         int err = 0;
1056         u64 last_start;
1057         u64 last_end;
1058
1059         btrfs_debug_check_extent_io_range(tree, start, end);
1060
1061 again:
1062         if (!prealloc && (mask & __GFP_WAIT)) {
1063                 prealloc = alloc_extent_state(mask);
1064                 if (!prealloc)
1065                         return -ENOMEM;
1066         }
1067
1068         spin_lock(&tree->lock);
1069         if (cached_state && *cached_state) {
1070                 state = *cached_state;
1071                 if (state->start <= start && state->end > start &&
1072                     state->tree) {
1073                         node = &state->rb_node;
1074                         goto hit_next;
1075                 }
1076         }
1077
1078         /*
1079          * this search will find all the extents that end after
1080          * our range starts.
1081          */
1082         node = tree_search_for_insert(tree, start, &p, &parent);
1083         if (!node) {
1084                 prealloc = alloc_extent_state_atomic(prealloc);
1085                 if (!prealloc) {
1086                         err = -ENOMEM;
1087                         goto out;
1088                 }
1089                 err = insert_state(tree, prealloc, start, end,
1090                                    &p, &parent, &bits);
1091                 if (err)
1092                         extent_io_tree_panic(tree, err);
1093                 cache_state(prealloc, cached_state);
1094                 prealloc = NULL;
1095                 goto out;
1096         }
1097         state = rb_entry(node, struct extent_state, rb_node);
1098 hit_next:
1099         last_start = state->start;
1100         last_end = state->end;
1101
1102         /*
1103          * | ---- desired range ---- |
1104          * | state |
1105          *
1106          * Just lock what we found and keep going
1107          */
1108         if (state->start == start && state->end <= end) {
1109                 set_state_bits(tree, state, &bits);
1110                 cache_state(state, cached_state);
1111                 state = clear_state_bit(tree, state, &clear_bits, 0);
1112                 if (last_end == (u64)-1)
1113                         goto out;
1114                 start = last_end + 1;
1115                 if (start < end && state && state->start == start &&
1116                     !need_resched())
1117                         goto hit_next;
1118                 goto search_again;
1119         }
1120
1121         /*
1122          *     | ---- desired range ---- |
1123          * | state |
1124          *   or
1125          * | ------------- state -------------- |
1126          *
1127          * We need to split the extent we found, and may flip bits on
1128          * second half.
1129          *
1130          * If the extent we found extends past our
1131          * range, we just split and search again.  It'll get split
1132          * again the next time though.
1133          *
1134          * If the extent we found is inside our range, we set the
1135          * desired bit on it.
1136          */
1137         if (state->start < start) {
1138                 prealloc = alloc_extent_state_atomic(prealloc);
1139                 if (!prealloc) {
1140                         err = -ENOMEM;
1141                         goto out;
1142                 }
1143                 err = split_state(tree, state, prealloc, start);
1144                 if (err)
1145                         extent_io_tree_panic(tree, err);
1146                 prealloc = NULL;
1147                 if (err)
1148                         goto out;
1149                 if (state->end <= end) {
1150                         set_state_bits(tree, state, &bits);
1151                         cache_state(state, cached_state);
1152                         state = clear_state_bit(tree, state, &clear_bits, 0);
1153                         if (last_end == (u64)-1)
1154                                 goto out;
1155                         start = last_end + 1;
1156                         if (start < end && state && state->start == start &&
1157                             !need_resched())
1158                                 goto hit_next;
1159                 }
1160                 goto search_again;
1161         }
1162         /*
1163          * | ---- desired range ---- |
1164          *     | state | or               | state |
1165          *
1166          * There's a hole, we need to insert something in it and
1167          * ignore the extent we found.
1168          */
1169         if (state->start > start) {
1170                 u64 this_end;
1171                 if (end < last_start)
1172                         this_end = end;
1173                 else
1174                         this_end = last_start - 1;
1175
1176                 prealloc = alloc_extent_state_atomic(prealloc);
1177                 if (!prealloc) {
1178                         err = -ENOMEM;
1179                         goto out;
1180                 }
1181
1182                 /*
1183                  * Avoid to free 'prealloc' if it can be merged with
1184                  * the later extent.
1185                  */
1186                 err = insert_state(tree, prealloc, start, this_end,
1187                                    NULL, NULL, &bits);
1188                 if (err)
1189                         extent_io_tree_panic(tree, err);
1190                 cache_state(prealloc, cached_state);
1191                 prealloc = NULL;
1192                 start = this_end + 1;
1193                 goto search_again;
1194         }
1195         /*
1196          * | ---- desired range ---- |
1197          *                        | state |
1198          * We need to split the extent, and set the bit
1199          * on the first half
1200          */
1201         if (state->start <= end && state->end > end) {
1202                 prealloc = alloc_extent_state_atomic(prealloc);
1203                 if (!prealloc) {
1204                         err = -ENOMEM;
1205                         goto out;
1206                 }
1207
1208                 err = split_state(tree, state, prealloc, end + 1);
1209                 if (err)
1210                         extent_io_tree_panic(tree, err);
1211
1212                 set_state_bits(tree, prealloc, &bits);
1213                 cache_state(prealloc, cached_state);
1214                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1215                 prealloc = NULL;
1216                 goto out;
1217         }
1218
1219         goto search_again;
1220
1221 out:
1222         spin_unlock(&tree->lock);
1223         if (prealloc)
1224                 free_extent_state(prealloc);
1225
1226         return err;
1227
1228 search_again:
1229         if (start > end)
1230                 goto out;
1231         spin_unlock(&tree->lock);
1232         if (mask & __GFP_WAIT)
1233                 cond_resched();
1234         goto again;
1235 }
1236
1237 /* wrappers around set/clear extent bit */
1238 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1239                      gfp_t mask)
1240 {
1241         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1242                               NULL, mask);
1243 }
1244
1245 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1246                     unsigned long bits, gfp_t mask)
1247 {
1248         return set_extent_bit(tree, start, end, bits, NULL,
1249                               NULL, mask);
1250 }
1251
1252 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1253                       unsigned long bits, gfp_t mask)
1254 {
1255         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1256 }
1257
1258 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1259                         struct extent_state **cached_state, gfp_t mask)
1260 {
1261         return set_extent_bit(tree, start, end,
1262                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1263                               NULL, cached_state, mask);
1264 }
1265
1266 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1267                       struct extent_state **cached_state, gfp_t mask)
1268 {
1269         return set_extent_bit(tree, start, end,
1270                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1271                               NULL, cached_state, mask);
1272 }
1273
1274 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1275                        gfp_t mask)
1276 {
1277         return clear_extent_bit(tree, start, end,
1278                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1279                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1280 }
1281
1282 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1283                      gfp_t mask)
1284 {
1285         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1286                               NULL, mask);
1287 }
1288
1289 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1290                         struct extent_state **cached_state, gfp_t mask)
1291 {
1292         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1293                               cached_state, mask);
1294 }
1295
1296 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1297                           struct extent_state **cached_state, gfp_t mask)
1298 {
1299         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1300                                 cached_state, mask);
1301 }
1302
1303 /*
1304  * either insert or lock state struct between start and end use mask to tell
1305  * us if waiting is desired.
1306  */
1307 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1308                      unsigned long bits, struct extent_state **cached_state)
1309 {
1310         int err;
1311         u64 failed_start;
1312         while (1) {
1313                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1314                                        EXTENT_LOCKED, &failed_start,
1315                                        cached_state, GFP_NOFS);
1316                 if (err == -EEXIST) {
1317                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1318                         start = failed_start;
1319                 } else
1320                         break;
1321                 WARN_ON(start > end);
1322         }
1323         return err;
1324 }
1325
1326 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1327 {
1328         return lock_extent_bits(tree, start, end, 0, NULL);
1329 }
1330
1331 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1332 {
1333         int err;
1334         u64 failed_start;
1335
1336         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1337                                &failed_start, NULL, GFP_NOFS);
1338         if (err == -EEXIST) {
1339                 if (failed_start > start)
1340                         clear_extent_bit(tree, start, failed_start - 1,
1341                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1342                 return 0;
1343         }
1344         return 1;
1345 }
1346
1347 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1348                          struct extent_state **cached, gfp_t mask)
1349 {
1350         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1351                                 mask);
1352 }
1353
1354 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1355 {
1356         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1357                                 GFP_NOFS);
1358 }
1359
1360 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1361 {
1362         unsigned long index = start >> PAGE_CACHE_SHIFT;
1363         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1364         struct page *page;
1365
1366         while (index <= end_index) {
1367                 page = find_get_page(inode->i_mapping, index);
1368                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1369                 clear_page_dirty_for_io(page);
1370                 page_cache_release(page);
1371                 index++;
1372         }
1373         return 0;
1374 }
1375
1376 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1377 {
1378         unsigned long index = start >> PAGE_CACHE_SHIFT;
1379         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1380         struct page *page;
1381
1382         while (index <= end_index) {
1383                 page = find_get_page(inode->i_mapping, index);
1384                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1385                 account_page_redirty(page);
1386                 __set_page_dirty_nobuffers(page);
1387                 page_cache_release(page);
1388                 index++;
1389         }
1390         return 0;
1391 }
1392
1393 /*
1394  * helper function to set both pages and extents in the tree writeback
1395  */
1396 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1397 {
1398         unsigned long index = start >> PAGE_CACHE_SHIFT;
1399         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1400         struct page *page;
1401
1402         while (index <= end_index) {
1403                 page = find_get_page(tree->mapping, index);
1404                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1405                 set_page_writeback(page);
1406                 page_cache_release(page);
1407                 index++;
1408         }
1409         return 0;
1410 }
1411
1412 /* find the first state struct with 'bits' set after 'start', and
1413  * return it.  tree->lock must be held.  NULL will returned if
1414  * nothing was found after 'start'
1415  */
1416 static struct extent_state *
1417 find_first_extent_bit_state(struct extent_io_tree *tree,
1418                             u64 start, unsigned long bits)
1419 {
1420         struct rb_node *node;
1421         struct extent_state *state;
1422
1423         /*
1424          * this search will find all the extents that end after
1425          * our range starts.
1426          */
1427         node = tree_search(tree, start);
1428         if (!node)
1429                 goto out;
1430
1431         while (1) {
1432                 state = rb_entry(node, struct extent_state, rb_node);
1433                 if (state->end >= start && (state->state & bits))
1434                         return state;
1435
1436                 node = rb_next(node);
1437                 if (!node)
1438                         break;
1439         }
1440 out:
1441         return NULL;
1442 }
1443
1444 /*
1445  * find the first offset in the io tree with 'bits' set. zero is
1446  * returned if we find something, and *start_ret and *end_ret are
1447  * set to reflect the state struct that was found.
1448  *
1449  * If nothing was found, 1 is returned. If found something, return 0.
1450  */
1451 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1452                           u64 *start_ret, u64 *end_ret, unsigned long bits,
1453                           struct extent_state **cached_state)
1454 {
1455         struct extent_state *state;
1456         struct rb_node *n;
1457         int ret = 1;
1458
1459         spin_lock(&tree->lock);
1460         if (cached_state && *cached_state) {
1461                 state = *cached_state;
1462                 if (state->end == start - 1 && state->tree) {
1463                         n = rb_next(&state->rb_node);
1464                         while (n) {
1465                                 state = rb_entry(n, struct extent_state,
1466                                                  rb_node);
1467                                 if (state->state & bits)
1468                                         goto got_it;
1469                                 n = rb_next(n);
1470                         }
1471                         free_extent_state(*cached_state);
1472                         *cached_state = NULL;
1473                         goto out;
1474                 }
1475                 free_extent_state(*cached_state);
1476                 *cached_state = NULL;
1477         }
1478
1479         state = find_first_extent_bit_state(tree, start, bits);
1480 got_it:
1481         if (state) {
1482                 cache_state(state, cached_state);
1483                 *start_ret = state->start;
1484                 *end_ret = state->end;
1485                 ret = 0;
1486         }
1487 out:
1488         spin_unlock(&tree->lock);
1489         return ret;
1490 }
1491
1492 /*
1493  * find a contiguous range of bytes in the file marked as delalloc, not
1494  * more than 'max_bytes'.  start and end are used to return the range,
1495  *
1496  * 1 is returned if we find something, 0 if nothing was in the tree
1497  */
1498 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1499                                         u64 *start, u64 *end, u64 max_bytes,
1500                                         struct extent_state **cached_state)
1501 {
1502         struct rb_node *node;
1503         struct extent_state *state;
1504         u64 cur_start = *start;
1505         u64 found = 0;
1506         u64 total_bytes = 0;
1507
1508         spin_lock(&tree->lock);
1509
1510         /*
1511          * this search will find all the extents that end after
1512          * our range starts.
1513          */
1514         node = tree_search(tree, cur_start);
1515         if (!node) {
1516                 if (!found)
1517                         *end = (u64)-1;
1518                 goto out;
1519         }
1520
1521         while (1) {
1522                 state = rb_entry(node, struct extent_state, rb_node);
1523                 if (found && (state->start != cur_start ||
1524                               (state->state & EXTENT_BOUNDARY))) {
1525                         goto out;
1526                 }
1527                 if (!(state->state & EXTENT_DELALLOC)) {
1528                         if (!found)
1529                                 *end = state->end;
1530                         goto out;
1531                 }
1532                 if (!found) {
1533                         *start = state->start;
1534                         *cached_state = state;
1535                         atomic_inc(&state->refs);
1536                 }
1537                 found++;
1538                 *end = state->end;
1539                 cur_start = state->end + 1;
1540                 node = rb_next(node);
1541                 total_bytes += state->end - state->start + 1;
1542                 if (total_bytes >= max_bytes)
1543                         break;
1544                 if (!node)
1545                         break;
1546         }
1547 out:
1548         spin_unlock(&tree->lock);
1549         return found;
1550 }
1551
1552 static noinline void __unlock_for_delalloc(struct inode *inode,
1553                                            struct page *locked_page,
1554                                            u64 start, u64 end)
1555 {
1556         int ret;
1557         struct page *pages[16];
1558         unsigned long index = start >> PAGE_CACHE_SHIFT;
1559         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1560         unsigned long nr_pages = end_index - index + 1;
1561         int i;
1562
1563         if (index == locked_page->index && end_index == index)
1564                 return;
1565
1566         while (nr_pages > 0) {
1567                 ret = find_get_pages_contig(inode->i_mapping, index,
1568                                      min_t(unsigned long, nr_pages,
1569                                      ARRAY_SIZE(pages)), pages);
1570                 for (i = 0; i < ret; i++) {
1571                         if (pages[i] != locked_page)
1572                                 unlock_page(pages[i]);
1573                         page_cache_release(pages[i]);
1574                 }
1575                 nr_pages -= ret;
1576                 index += ret;
1577                 cond_resched();
1578         }
1579 }
1580
1581 static noinline int lock_delalloc_pages(struct inode *inode,
1582                                         struct page *locked_page,
1583                                         u64 delalloc_start,
1584                                         u64 delalloc_end)
1585 {
1586         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1587         unsigned long start_index = index;
1588         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1589         unsigned long pages_locked = 0;
1590         struct page *pages[16];
1591         unsigned long nrpages;
1592         int ret;
1593         int i;
1594
1595         /* the caller is responsible for locking the start index */
1596         if (index == locked_page->index && index == end_index)
1597                 return 0;
1598
1599         /* skip the page at the start index */
1600         nrpages = end_index - index + 1;
1601         while (nrpages > 0) {
1602                 ret = find_get_pages_contig(inode->i_mapping, index,
1603                                      min_t(unsigned long,
1604                                      nrpages, ARRAY_SIZE(pages)), pages);
1605                 if (ret == 0) {
1606                         ret = -EAGAIN;
1607                         goto done;
1608                 }
1609                 /* now we have an array of pages, lock them all */
1610                 for (i = 0; i < ret; i++) {
1611                         /*
1612                          * the caller is taking responsibility for
1613                          * locked_page
1614                          */
1615                         if (pages[i] != locked_page) {
1616                                 lock_page(pages[i]);
1617                                 if (!PageDirty(pages[i]) ||
1618                                     pages[i]->mapping != inode->i_mapping) {
1619                                         ret = -EAGAIN;
1620                                         unlock_page(pages[i]);
1621                                         page_cache_release(pages[i]);
1622                                         goto done;
1623                                 }
1624                         }
1625                         page_cache_release(pages[i]);
1626                         pages_locked++;
1627                 }
1628                 nrpages -= ret;
1629                 index += ret;
1630                 cond_resched();
1631         }
1632         ret = 0;
1633 done:
1634         if (ret && pages_locked) {
1635                 __unlock_for_delalloc(inode, locked_page,
1636                               delalloc_start,
1637                               ((u64)(start_index + pages_locked - 1)) <<
1638                               PAGE_CACHE_SHIFT);
1639         }
1640         return ret;
1641 }
1642
1643 /*
1644  * find a contiguous range of bytes in the file marked as delalloc, not
1645  * more than 'max_bytes'.  start and end are used to return the range,
1646  *
1647  * 1 is returned if we find something, 0 if nothing was in the tree
1648  */
1649 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1650                                     struct extent_io_tree *tree,
1651                                     struct page *locked_page, u64 *start,
1652                                     u64 *end, u64 max_bytes)
1653 {
1654         u64 delalloc_start;
1655         u64 delalloc_end;
1656         u64 found;
1657         struct extent_state *cached_state = NULL;
1658         int ret;
1659         int loops = 0;
1660
1661 again:
1662         /* step one, find a bunch of delalloc bytes starting at start */
1663         delalloc_start = *start;
1664         delalloc_end = 0;
1665         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1666                                     max_bytes, &cached_state);
1667         if (!found || delalloc_end <= *start) {
1668                 *start = delalloc_start;
1669                 *end = delalloc_end;
1670                 free_extent_state(cached_state);
1671                 return 0;
1672         }
1673
1674         /*
1675          * start comes from the offset of locked_page.  We have to lock
1676          * pages in order, so we can't process delalloc bytes before
1677          * locked_page
1678          */
1679         if (delalloc_start < *start)
1680                 delalloc_start = *start;
1681
1682         /*
1683          * make sure to limit the number of pages we try to lock down
1684          */
1685         if (delalloc_end + 1 - delalloc_start > max_bytes)
1686                 delalloc_end = delalloc_start + max_bytes - 1;
1687
1688         /* step two, lock all the pages after the page that has start */
1689         ret = lock_delalloc_pages(inode, locked_page,
1690                                   delalloc_start, delalloc_end);
1691         if (ret == -EAGAIN) {
1692                 /* some of the pages are gone, lets avoid looping by
1693                  * shortening the size of the delalloc range we're searching
1694                  */
1695                 free_extent_state(cached_state);
1696                 if (!loops) {
1697                         max_bytes = PAGE_CACHE_SIZE;
1698                         loops = 1;
1699                         goto again;
1700                 } else {
1701                         found = 0;
1702                         goto out_failed;
1703                 }
1704         }
1705         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1706
1707         /* step three, lock the state bits for the whole range */
1708         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1709
1710         /* then test to make sure it is all still delalloc */
1711         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1712                              EXTENT_DELALLOC, 1, cached_state);
1713         if (!ret) {
1714                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1715                                      &cached_state, GFP_NOFS);
1716                 __unlock_for_delalloc(inode, locked_page,
1717                               delalloc_start, delalloc_end);
1718                 cond_resched();
1719                 goto again;
1720         }
1721         free_extent_state(cached_state);
1722         *start = delalloc_start;
1723         *end = delalloc_end;
1724 out_failed:
1725         return found;
1726 }
1727
1728 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1729                                  struct page *locked_page,
1730                                  unsigned long clear_bits,
1731                                  unsigned long page_ops)
1732 {
1733         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1734         int ret;
1735         struct page *pages[16];
1736         unsigned long index = start >> PAGE_CACHE_SHIFT;
1737         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1738         unsigned long nr_pages = end_index - index + 1;
1739         int i;
1740
1741         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1742         if (page_ops == 0)
1743                 return 0;
1744
1745         while (nr_pages > 0) {
1746                 ret = find_get_pages_contig(inode->i_mapping, index,
1747                                      min_t(unsigned long,
1748                                      nr_pages, ARRAY_SIZE(pages)), pages);
1749                 for (i = 0; i < ret; i++) {
1750
1751                         if (page_ops & PAGE_SET_PRIVATE2)
1752                                 SetPagePrivate2(pages[i]);
1753
1754                         if (pages[i] == locked_page) {
1755                                 page_cache_release(pages[i]);
1756                                 continue;
1757                         }
1758                         if (page_ops & PAGE_CLEAR_DIRTY)
1759                                 clear_page_dirty_for_io(pages[i]);
1760                         if (page_ops & PAGE_SET_WRITEBACK)
1761                                 set_page_writeback(pages[i]);
1762                         if (page_ops & PAGE_END_WRITEBACK)
1763                                 end_page_writeback(pages[i]);
1764                         if (page_ops & PAGE_UNLOCK)
1765                                 unlock_page(pages[i]);
1766                         page_cache_release(pages[i]);
1767                 }
1768                 nr_pages -= ret;
1769                 index += ret;
1770                 cond_resched();
1771         }
1772         return 0;
1773 }
1774
1775 /*
1776  * count the number of bytes in the tree that have a given bit(s)
1777  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1778  * cached.  The total number found is returned.
1779  */
1780 u64 count_range_bits(struct extent_io_tree *tree,
1781                      u64 *start, u64 search_end, u64 max_bytes,
1782                      unsigned long bits, int contig)
1783 {
1784         struct rb_node *node;
1785         struct extent_state *state;
1786         u64 cur_start = *start;
1787         u64 total_bytes = 0;
1788         u64 last = 0;
1789         int found = 0;
1790
1791         if (WARN_ON(search_end <= cur_start))
1792                 return 0;
1793
1794         spin_lock(&tree->lock);
1795         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1796                 total_bytes = tree->dirty_bytes;
1797                 goto out;
1798         }
1799         /*
1800          * this search will find all the extents that end after
1801          * our range starts.
1802          */
1803         node = tree_search(tree, cur_start);
1804         if (!node)
1805                 goto out;
1806
1807         while (1) {
1808                 state = rb_entry(node, struct extent_state, rb_node);
1809                 if (state->start > search_end)
1810                         break;
1811                 if (contig && found && state->start > last + 1)
1812                         break;
1813                 if (state->end >= cur_start && (state->state & bits) == bits) {
1814                         total_bytes += min(search_end, state->end) + 1 -
1815                                        max(cur_start, state->start);
1816                         if (total_bytes >= max_bytes)
1817                                 break;
1818                         if (!found) {
1819                                 *start = max(cur_start, state->start);
1820                                 found = 1;
1821                         }
1822                         last = state->end;
1823                 } else if (contig && found) {
1824                         break;
1825                 }
1826                 node = rb_next(node);
1827                 if (!node)
1828                         break;
1829         }
1830 out:
1831         spin_unlock(&tree->lock);
1832         return total_bytes;
1833 }
1834
1835 /*
1836  * set the private field for a given byte offset in the tree.  If there isn't
1837  * an extent_state there already, this does nothing.
1838  */
1839 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1840 {
1841         struct rb_node *node;
1842         struct extent_state *state;
1843         int ret = 0;
1844
1845         spin_lock(&tree->lock);
1846         /*
1847          * this search will find all the extents that end after
1848          * our range starts.
1849          */
1850         node = tree_search(tree, start);
1851         if (!node) {
1852                 ret = -ENOENT;
1853                 goto out;
1854         }
1855         state = rb_entry(node, struct extent_state, rb_node);
1856         if (state->start != start) {
1857                 ret = -ENOENT;
1858                 goto out;
1859         }
1860         state->private = private;
1861 out:
1862         spin_unlock(&tree->lock);
1863         return ret;
1864 }
1865
1866 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1867 {
1868         struct rb_node *node;
1869         struct extent_state *state;
1870         int ret = 0;
1871
1872         spin_lock(&tree->lock);
1873         /*
1874          * this search will find all the extents that end after
1875          * our range starts.
1876          */
1877         node = tree_search(tree, start);
1878         if (!node) {
1879                 ret = -ENOENT;
1880                 goto out;
1881         }
1882         state = rb_entry(node, struct extent_state, rb_node);
1883         if (state->start != start) {
1884                 ret = -ENOENT;
1885                 goto out;
1886         }
1887         *private = state->private;
1888 out:
1889         spin_unlock(&tree->lock);
1890         return ret;
1891 }
1892
1893 /*
1894  * searches a range in the state tree for a given mask.
1895  * If 'filled' == 1, this returns 1 only if every extent in the tree
1896  * has the bits set.  Otherwise, 1 is returned if any bit in the
1897  * range is found set.
1898  */
1899 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1900                    unsigned long bits, int filled, struct extent_state *cached)
1901 {
1902         struct extent_state *state = NULL;
1903         struct rb_node *node;
1904         int bitset = 0;
1905
1906         spin_lock(&tree->lock);
1907         if (cached && cached->tree && cached->start <= start &&
1908             cached->end > start)
1909                 node = &cached->rb_node;
1910         else
1911                 node = tree_search(tree, start);
1912         while (node && start <= end) {
1913                 state = rb_entry(node, struct extent_state, rb_node);
1914
1915                 if (filled && state->start > start) {
1916                         bitset = 0;
1917                         break;
1918                 }
1919
1920                 if (state->start > end)
1921                         break;
1922
1923                 if (state->state & bits) {
1924                         bitset = 1;
1925                         if (!filled)
1926                                 break;
1927                 } else if (filled) {
1928                         bitset = 0;
1929                         break;
1930                 }
1931
1932                 if (state->end == (u64)-1)
1933                         break;
1934
1935                 start = state->end + 1;
1936                 if (start > end)
1937                         break;
1938                 node = rb_next(node);
1939                 if (!node) {
1940                         if (filled)
1941                                 bitset = 0;
1942                         break;
1943                 }
1944         }
1945         spin_unlock(&tree->lock);
1946         return bitset;
1947 }
1948
1949 /*
1950  * helper function to set a given page up to date if all the
1951  * extents in the tree for that page are up to date
1952  */
1953 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1954 {
1955         u64 start = page_offset(page);
1956         u64 end = start + PAGE_CACHE_SIZE - 1;
1957         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1958                 SetPageUptodate(page);
1959 }
1960
1961 /*
1962  * When IO fails, either with EIO or csum verification fails, we
1963  * try other mirrors that might have a good copy of the data.  This
1964  * io_failure_record is used to record state as we go through all the
1965  * mirrors.  If another mirror has good data, the page is set up to date
1966  * and things continue.  If a good mirror can't be found, the original
1967  * bio end_io callback is called to indicate things have failed.
1968  */
1969 struct io_failure_record {
1970         struct page *page;
1971         u64 start;
1972         u64 len;
1973         u64 logical;
1974         unsigned long bio_flags;
1975         int this_mirror;
1976         int failed_mirror;
1977         int in_validation;
1978 };
1979
1980 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1981                                 int did_repair)
1982 {
1983         int ret;
1984         int err = 0;
1985         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1986
1987         set_state_private(failure_tree, rec->start, 0);
1988         ret = clear_extent_bits(failure_tree, rec->start,
1989                                 rec->start + rec->len - 1,
1990                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1991         if (ret)
1992                 err = ret;
1993
1994         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1995                                 rec->start + rec->len - 1,
1996                                 EXTENT_DAMAGED, GFP_NOFS);
1997         if (ret && !err)
1998                 err = ret;
1999
2000         kfree(rec);
2001         return err;
2002 }
2003
2004 /*
2005  * this bypasses the standard btrfs submit functions deliberately, as
2006  * the standard behavior is to write all copies in a raid setup. here we only
2007  * want to write the one bad copy. so we do the mapping for ourselves and issue
2008  * submit_bio directly.
2009  * to avoid any synchronization issues, wait for the data after writing, which
2010  * actually prevents the read that triggered the error from finishing.
2011  * currently, there can be no more than two copies of every data bit. thus,
2012  * exactly one rewrite is required.
2013  */
2014 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 start,
2015                         u64 length, u64 logical, struct page *page,
2016                         int mirror_num)
2017 {
2018         struct bio *bio;
2019         struct btrfs_device *dev;
2020         u64 map_length = 0;
2021         u64 sector;
2022         struct btrfs_bio *bbio = NULL;
2023         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2024         int ret;
2025
2026         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2027         BUG_ON(!mirror_num);
2028
2029         /* we can't repair anything in raid56 yet */
2030         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2031                 return 0;
2032
2033         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2034         if (!bio)
2035                 return -EIO;
2036         bio->bi_iter.bi_size = 0;
2037         map_length = length;
2038
2039         ret = btrfs_map_block(fs_info, WRITE, logical,
2040                               &map_length, &bbio, mirror_num);
2041         if (ret) {
2042                 bio_put(bio);
2043                 return -EIO;
2044         }
2045         BUG_ON(mirror_num != bbio->mirror_num);
2046         sector = bbio->stripes[mirror_num-1].physical >> 9;
2047         bio->bi_iter.bi_sector = sector;
2048         dev = bbio->stripes[mirror_num-1].dev;
2049         kfree(bbio);
2050         if (!dev || !dev->bdev || !dev->writeable) {
2051                 bio_put(bio);
2052                 return -EIO;
2053         }
2054         bio->bi_bdev = dev->bdev;
2055         bio_add_page(bio, page, length, start - page_offset(page));
2056
2057         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio)) {
2058                 /* try to remap that extent elsewhere? */
2059                 bio_put(bio);
2060                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2061                 return -EIO;
2062         }
2063
2064         printk_ratelimited_in_rcu(KERN_INFO
2065                         "BTRFS: read error corrected: ino %lu off %llu "
2066                     "(dev %s sector %llu)\n", page->mapping->host->i_ino,
2067                     start, rcu_str_deref(dev->name), sector);
2068
2069         bio_put(bio);
2070         return 0;
2071 }
2072
2073 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2074                          int mirror_num)
2075 {
2076         u64 start = eb->start;
2077         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2078         int ret = 0;
2079
2080         if (root->fs_info->sb->s_flags & MS_RDONLY)
2081                 return -EROFS;
2082
2083         for (i = 0; i < num_pages; i++) {
2084                 struct page *p = extent_buffer_page(eb, i);
2085                 ret = repair_io_failure(root->fs_info, start, PAGE_CACHE_SIZE,
2086                                         start, p, mirror_num);
2087                 if (ret)
2088                         break;
2089                 start += PAGE_CACHE_SIZE;
2090         }
2091
2092         return ret;
2093 }
2094
2095 /*
2096  * each time an IO finishes, we do a fast check in the IO failure tree
2097  * to see if we need to process or clean up an io_failure_record
2098  */
2099 static int clean_io_failure(u64 start, struct page *page)
2100 {
2101         u64 private;
2102         u64 private_failure;
2103         struct io_failure_record *failrec;
2104         struct inode *inode = page->mapping->host;
2105         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2106         struct extent_state *state;
2107         int num_copies;
2108         int did_repair = 0;
2109         int ret;
2110
2111         private = 0;
2112         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2113                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2114         if (!ret)
2115                 return 0;
2116
2117         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2118                                 &private_failure);
2119         if (ret)
2120                 return 0;
2121
2122         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2123         BUG_ON(!failrec->this_mirror);
2124
2125         if (failrec->in_validation) {
2126                 /* there was no real error, just free the record */
2127                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2128                          failrec->start);
2129                 did_repair = 1;
2130                 goto out;
2131         }
2132         if (fs_info->sb->s_flags & MS_RDONLY)
2133                 goto out;
2134
2135         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2136         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2137                                             failrec->start,
2138                                             EXTENT_LOCKED);
2139         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2140
2141         if (state && state->start <= failrec->start &&
2142             state->end >= failrec->start + failrec->len - 1) {
2143                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2144                                               failrec->len);
2145                 if (num_copies > 1)  {
2146                         ret = repair_io_failure(fs_info, start, failrec->len,
2147                                                 failrec->logical, page,
2148                                                 failrec->failed_mirror);
2149                         did_repair = !ret;
2150                 }
2151                 ret = 0;
2152         }
2153
2154 out:
2155         if (!ret)
2156                 ret = free_io_failure(inode, failrec, did_repair);
2157
2158         return ret;
2159 }
2160
2161 /*
2162  * this is a generic handler for readpage errors (default
2163  * readpage_io_failed_hook). if other copies exist, read those and write back
2164  * good data to the failed position. does not investigate in remapping the
2165  * failed extent elsewhere, hoping the device will be smart enough to do this as
2166  * needed
2167  */
2168
2169 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2170                               struct page *page, u64 start, u64 end,
2171                               int failed_mirror)
2172 {
2173         struct io_failure_record *failrec = NULL;
2174         u64 private;
2175         struct extent_map *em;
2176         struct inode *inode = page->mapping->host;
2177         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2178         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2179         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2180         struct bio *bio;
2181         struct btrfs_io_bio *btrfs_failed_bio;
2182         struct btrfs_io_bio *btrfs_bio;
2183         int num_copies;
2184         int ret;
2185         int read_mode;
2186         u64 logical;
2187
2188         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2189
2190         ret = get_state_private(failure_tree, start, &private);
2191         if (ret) {
2192                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2193                 if (!failrec)
2194                         return -ENOMEM;
2195                 failrec->start = start;
2196                 failrec->len = end - start + 1;
2197                 failrec->this_mirror = 0;
2198                 failrec->bio_flags = 0;
2199                 failrec->in_validation = 0;
2200
2201                 read_lock(&em_tree->lock);
2202                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2203                 if (!em) {
2204                         read_unlock(&em_tree->lock);
2205                         kfree(failrec);
2206                         return -EIO;
2207                 }
2208
2209                 if (em->start > start || em->start + em->len <= start) {
2210                         free_extent_map(em);
2211                         em = NULL;
2212                 }
2213                 read_unlock(&em_tree->lock);
2214
2215                 if (!em) {
2216                         kfree(failrec);
2217                         return -EIO;
2218                 }
2219                 logical = start - em->start;
2220                 logical = em->block_start + logical;
2221                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2222                         logical = em->block_start;
2223                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2224                         extent_set_compress_type(&failrec->bio_flags,
2225                                                  em->compress_type);
2226                 }
2227                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2228                          "len=%llu\n", logical, start, failrec->len);
2229                 failrec->logical = logical;
2230                 free_extent_map(em);
2231
2232                 /* set the bits in the private failure tree */
2233                 ret = set_extent_bits(failure_tree, start, end,
2234                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2235                 if (ret >= 0)
2236                         ret = set_state_private(failure_tree, start,
2237                                                 (u64)(unsigned long)failrec);
2238                 /* set the bits in the inode's tree */
2239                 if (ret >= 0)
2240                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2241                                                 GFP_NOFS);
2242                 if (ret < 0) {
2243                         kfree(failrec);
2244                         return ret;
2245                 }
2246         } else {
2247                 failrec = (struct io_failure_record *)(unsigned long)private;
2248                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2249                          "start=%llu, len=%llu, validation=%d\n",
2250                          failrec->logical, failrec->start, failrec->len,
2251                          failrec->in_validation);
2252                 /*
2253                  * when data can be on disk more than twice, add to failrec here
2254                  * (e.g. with a list for failed_mirror) to make
2255                  * clean_io_failure() clean all those errors at once.
2256                  */
2257         }
2258         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2259                                       failrec->logical, failrec->len);
2260         if (num_copies == 1) {
2261                 /*
2262                  * we only have a single copy of the data, so don't bother with
2263                  * all the retry and error correction code that follows. no
2264                  * matter what the error is, it is very likely to persist.
2265                  */
2266                 pr_debug("bio_readpage_error: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2267                          num_copies, failrec->this_mirror, failed_mirror);
2268                 free_io_failure(inode, failrec, 0);
2269                 return -EIO;
2270         }
2271
2272         /*
2273          * there are two premises:
2274          *      a) deliver good data to the caller
2275          *      b) correct the bad sectors on disk
2276          */
2277         if (failed_bio->bi_vcnt > 1) {
2278                 /*
2279                  * to fulfill b), we need to know the exact failing sectors, as
2280                  * we don't want to rewrite any more than the failed ones. thus,
2281                  * we need separate read requests for the failed bio
2282                  *
2283                  * if the following BUG_ON triggers, our validation request got
2284                  * merged. we need separate requests for our algorithm to work.
2285                  */
2286                 BUG_ON(failrec->in_validation);
2287                 failrec->in_validation = 1;
2288                 failrec->this_mirror = failed_mirror;
2289                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2290         } else {
2291                 /*
2292                  * we're ready to fulfill a) and b) alongside. get a good copy
2293                  * of the failed sector and if we succeed, we have setup
2294                  * everything for repair_io_failure to do the rest for us.
2295                  */
2296                 if (failrec->in_validation) {
2297                         BUG_ON(failrec->this_mirror != failed_mirror);
2298                         failrec->in_validation = 0;
2299                         failrec->this_mirror = 0;
2300                 }
2301                 failrec->failed_mirror = failed_mirror;
2302                 failrec->this_mirror++;
2303                 if (failrec->this_mirror == failed_mirror)
2304                         failrec->this_mirror++;
2305                 read_mode = READ_SYNC;
2306         }
2307
2308         if (failrec->this_mirror > num_copies) {
2309                 pr_debug("bio_readpage_error: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2310                          num_copies, failrec->this_mirror, failed_mirror);
2311                 free_io_failure(inode, failrec, 0);
2312                 return -EIO;
2313         }
2314
2315         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2316         if (!bio) {
2317                 free_io_failure(inode, failrec, 0);
2318                 return -EIO;
2319         }
2320         bio->bi_end_io = failed_bio->bi_end_io;
2321         bio->bi_iter.bi_sector = failrec->logical >> 9;
2322         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2323         bio->bi_iter.bi_size = 0;
2324
2325         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2326         if (btrfs_failed_bio->csum) {
2327                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2328                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2329
2330                 btrfs_bio = btrfs_io_bio(bio);
2331                 btrfs_bio->csum = btrfs_bio->csum_inline;
2332                 phy_offset >>= inode->i_sb->s_blocksize_bits;
2333                 phy_offset *= csum_size;
2334                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + phy_offset,
2335                        csum_size);
2336         }
2337
2338         bio_add_page(bio, page, failrec->len, start - page_offset(page));
2339
2340         pr_debug("bio_readpage_error: submitting new read[%#x] to "
2341                  "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2342                  failrec->this_mirror, num_copies, failrec->in_validation);
2343
2344         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2345                                          failrec->this_mirror,
2346                                          failrec->bio_flags, 0);
2347         return ret;
2348 }
2349
2350 /* lots and lots of room for performance fixes in the end_bio funcs */
2351
2352 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2353 {
2354         int uptodate = (err == 0);
2355         struct extent_io_tree *tree;
2356         int ret;
2357
2358         tree = &BTRFS_I(page->mapping->host)->io_tree;
2359
2360         if (tree->ops && tree->ops->writepage_end_io_hook) {
2361                 ret = tree->ops->writepage_end_io_hook(page, start,
2362                                                end, NULL, uptodate);
2363                 if (ret)
2364                         uptodate = 0;
2365         }
2366
2367         if (!uptodate) {
2368                 ClearPageUptodate(page);
2369                 SetPageError(page);
2370                 ret = ret < 0 ? ret : -EIO;
2371                 mapping_set_error(page->mapping, ret);
2372         }
2373         return 0;
2374 }
2375
2376 /*
2377  * after a writepage IO is done, we need to:
2378  * clear the uptodate bits on error
2379  * clear the writeback bits in the extent tree for this IO
2380  * end_page_writeback if the page has no more pending IO
2381  *
2382  * Scheduling is not allowed, so the extent state tree is expected
2383  * to have one and only one object corresponding to this IO.
2384  */
2385 static void end_bio_extent_writepage(struct bio *bio, int err)
2386 {
2387         struct bio_vec *bvec;
2388         u64 start;
2389         u64 end;
2390         int i;
2391
2392         bio_for_each_segment_all(bvec, bio, i) {
2393                 struct page *page = bvec->bv_page;
2394
2395                 /* We always issue full-page reads, but if some block
2396                  * in a page fails to read, blk_update_request() will
2397                  * advance bv_offset and adjust bv_len to compensate.
2398                  * Print a warning for nonzero offsets, and an error
2399                  * if they don't add up to a full page.  */
2400                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2401                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2402                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2403                                    "partial page write in btrfs with offset %u and length %u",
2404                                         bvec->bv_offset, bvec->bv_len);
2405                         else
2406                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2407                                    "incomplete page write in btrfs with offset %u and "
2408                                    "length %u",
2409                                         bvec->bv_offset, bvec->bv_len);
2410                 }
2411
2412                 start = page_offset(page);
2413                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2414
2415                 if (end_extent_writepage(page, err, start, end))
2416                         continue;
2417
2418                 end_page_writeback(page);
2419         }
2420
2421         bio_put(bio);
2422 }
2423
2424 static void
2425 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2426                               int uptodate)
2427 {
2428         struct extent_state *cached = NULL;
2429         u64 end = start + len - 1;
2430
2431         if (uptodate && tree->track_uptodate)
2432                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2433         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2434 }
2435
2436 /*
2437  * after a readpage IO is done, we need to:
2438  * clear the uptodate bits on error
2439  * set the uptodate bits if things worked
2440  * set the page up to date if all extents in the tree are uptodate
2441  * clear the lock bit in the extent tree
2442  * unlock the page if there are no other extents locked for it
2443  *
2444  * Scheduling is not allowed, so the extent state tree is expected
2445  * to have one and only one object corresponding to this IO.
2446  */
2447 static void end_bio_extent_readpage(struct bio *bio, int err)
2448 {
2449         struct bio_vec *bvec;
2450         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2451         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2452         struct extent_io_tree *tree;
2453         u64 offset = 0;
2454         u64 start;
2455         u64 end;
2456         u64 len;
2457         u64 extent_start = 0;
2458         u64 extent_len = 0;
2459         int mirror;
2460         int ret;
2461         int i;
2462
2463         if (err)
2464                 uptodate = 0;
2465
2466         bio_for_each_segment_all(bvec, bio, i) {
2467                 struct page *page = bvec->bv_page;
2468                 struct inode *inode = page->mapping->host;
2469
2470                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2471                          "mirror=%lu\n", (u64)bio->bi_iter.bi_sector, err,
2472                          io_bio->mirror_num);
2473                 tree = &BTRFS_I(inode)->io_tree;
2474
2475                 /* We always issue full-page reads, but if some block
2476                  * in a page fails to read, blk_update_request() will
2477                  * advance bv_offset and adjust bv_len to compensate.
2478                  * Print a warning for nonzero offsets, and an error
2479                  * if they don't add up to a full page.  */
2480                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2481                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2482                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2483                                    "partial page read in btrfs with offset %u and length %u",
2484                                         bvec->bv_offset, bvec->bv_len);
2485                         else
2486                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2487                                    "incomplete page read in btrfs with offset %u and "
2488                                    "length %u",
2489                                         bvec->bv_offset, bvec->bv_len);
2490                 }
2491
2492                 start = page_offset(page);
2493                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2494                 len = bvec->bv_len;
2495
2496                 mirror = io_bio->mirror_num;
2497                 if (likely(uptodate && tree->ops &&
2498                            tree->ops->readpage_end_io_hook)) {
2499                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2500                                                               page, start, end,
2501                                                               mirror);
2502                         if (ret)
2503                                 uptodate = 0;
2504                         else
2505                                 clean_io_failure(start, page);
2506                 }
2507
2508                 if (likely(uptodate))
2509                         goto readpage_ok;
2510
2511                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2512                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2513                         if (!ret && !err &&
2514                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2515                                 uptodate = 1;
2516                 } else {
2517                         /*
2518                          * The generic bio_readpage_error handles errors the
2519                          * following way: If possible, new read requests are
2520                          * created and submitted and will end up in
2521                          * end_bio_extent_readpage as well (if we're lucky, not
2522                          * in the !uptodate case). In that case it returns 0 and
2523                          * we just go on with the next page in our bio. If it
2524                          * can't handle the error it will return -EIO and we
2525                          * remain responsible for that page.
2526                          */
2527                         ret = bio_readpage_error(bio, offset, page, start, end,
2528                                                  mirror);
2529                         if (ret == 0) {
2530                                 uptodate =
2531                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2532                                 if (err)
2533                                         uptodate = 0;
2534                                 continue;
2535                         }
2536                 }
2537 readpage_ok:
2538                 if (likely(uptodate)) {
2539                         loff_t i_size = i_size_read(inode);
2540                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2541                         unsigned offset;
2542
2543                         /* Zero out the end if this page straddles i_size */
2544                         offset = i_size & (PAGE_CACHE_SIZE-1);
2545                         if (page->index == end_index && offset)
2546                                 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2547                         SetPageUptodate(page);
2548                 } else {
2549                         ClearPageUptodate(page);
2550                         SetPageError(page);
2551                 }
2552                 unlock_page(page);
2553                 offset += len;
2554
2555                 if (unlikely(!uptodate)) {
2556                         if (extent_len) {
2557                                 endio_readpage_release_extent(tree,
2558                                                               extent_start,
2559                                                               extent_len, 1);
2560                                 extent_start = 0;
2561                                 extent_len = 0;
2562                         }
2563                         endio_readpage_release_extent(tree, start,
2564                                                       end - start + 1, 0);
2565                 } else if (!extent_len) {
2566                         extent_start = start;
2567                         extent_len = end + 1 - start;
2568                 } else if (extent_start + extent_len == start) {
2569                         extent_len += end + 1 - start;
2570                 } else {
2571                         endio_readpage_release_extent(tree, extent_start,
2572                                                       extent_len, uptodate);
2573                         extent_start = start;
2574                         extent_len = end + 1 - start;
2575                 }
2576         }
2577
2578         if (extent_len)
2579                 endio_readpage_release_extent(tree, extent_start, extent_len,
2580                                               uptodate);
2581         if (io_bio->end_io)
2582                 io_bio->end_io(io_bio, err);
2583         bio_put(bio);
2584 }
2585
2586 /*
2587  * this allocates from the btrfs_bioset.  We're returning a bio right now
2588  * but you can call btrfs_io_bio for the appropriate container_of magic
2589  */
2590 struct bio *
2591 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2592                 gfp_t gfp_flags)
2593 {
2594         struct btrfs_io_bio *btrfs_bio;
2595         struct bio *bio;
2596
2597         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2598
2599         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2600                 while (!bio && (nr_vecs /= 2)) {
2601                         bio = bio_alloc_bioset(gfp_flags,
2602                                                nr_vecs, btrfs_bioset);
2603                 }
2604         }
2605
2606         if (bio) {
2607                 bio->bi_bdev = bdev;
2608                 bio->bi_iter.bi_sector = first_sector;
2609                 btrfs_bio = btrfs_io_bio(bio);
2610                 btrfs_bio->csum = NULL;
2611                 btrfs_bio->csum_allocated = NULL;
2612                 btrfs_bio->end_io = NULL;
2613         }
2614         return bio;
2615 }
2616
2617 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2618 {
2619         return bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2620 }
2621
2622
2623 /* this also allocates from the btrfs_bioset */
2624 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2625 {
2626         struct btrfs_io_bio *btrfs_bio;
2627         struct bio *bio;
2628
2629         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2630         if (bio) {
2631                 btrfs_bio = btrfs_io_bio(bio);
2632                 btrfs_bio->csum = NULL;
2633                 btrfs_bio->csum_allocated = NULL;
2634                 btrfs_bio->end_io = NULL;
2635         }
2636         return bio;
2637 }
2638
2639
2640 static int __must_check submit_one_bio(int rw, struct bio *bio,
2641                                        int mirror_num, unsigned long bio_flags)
2642 {
2643         int ret = 0;
2644         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2645         struct page *page = bvec->bv_page;
2646         struct extent_io_tree *tree = bio->bi_private;
2647         u64 start;
2648
2649         start = page_offset(page) + bvec->bv_offset;
2650
2651         bio->bi_private = NULL;
2652
2653         bio_get(bio);
2654
2655         if (tree->ops && tree->ops->submit_bio_hook)
2656                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2657                                            mirror_num, bio_flags, start);
2658         else
2659                 btrfsic_submit_bio(rw, bio);
2660
2661         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2662                 ret = -EOPNOTSUPP;
2663         bio_put(bio);
2664         return ret;
2665 }
2666
2667 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2668                      unsigned long offset, size_t size, struct bio *bio,
2669                      unsigned long bio_flags)
2670 {
2671         int ret = 0;
2672         if (tree->ops && tree->ops->merge_bio_hook)
2673                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2674                                                 bio_flags);
2675         BUG_ON(ret < 0);
2676         return ret;
2677
2678 }
2679
2680 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2681                               struct page *page, sector_t sector,
2682                               size_t size, unsigned long offset,
2683                               struct block_device *bdev,
2684                               struct bio **bio_ret,
2685                               unsigned long max_pages,
2686                               bio_end_io_t end_io_func,
2687                               int mirror_num,
2688                               unsigned long prev_bio_flags,
2689                               unsigned long bio_flags)
2690 {
2691         int ret = 0;
2692         struct bio *bio;
2693         int nr;
2694         int contig = 0;
2695         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2696         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2697         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2698
2699         if (bio_ret && *bio_ret) {
2700                 bio = *bio_ret;
2701                 if (old_compressed)
2702                         contig = bio->bi_iter.bi_sector == sector;
2703                 else
2704                         contig = bio_end_sector(bio) == sector;
2705
2706                 if (prev_bio_flags != bio_flags || !contig ||
2707                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2708                     bio_add_page(bio, page, page_size, offset) < page_size) {
2709                         ret = submit_one_bio(rw, bio, mirror_num,
2710                                              prev_bio_flags);
2711                         if (ret < 0)
2712                                 return ret;
2713                         bio = NULL;
2714                 } else {
2715                         return 0;
2716                 }
2717         }
2718         if (this_compressed)
2719                 nr = BIO_MAX_PAGES;
2720         else
2721                 nr = bio_get_nr_vecs(bdev);
2722
2723         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2724         if (!bio)
2725                 return -ENOMEM;
2726
2727         bio_add_page(bio, page, page_size, offset);
2728         bio->bi_end_io = end_io_func;
2729         bio->bi_private = tree;
2730
2731         if (bio_ret)
2732                 *bio_ret = bio;
2733         else
2734                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2735
2736         return ret;
2737 }
2738
2739 static void attach_extent_buffer_page(struct extent_buffer *eb,
2740                                       struct page *page)
2741 {
2742         if (!PagePrivate(page)) {
2743                 SetPagePrivate(page);
2744                 page_cache_get(page);
2745                 set_page_private(page, (unsigned long)eb);
2746         } else {
2747                 WARN_ON(page->private != (unsigned long)eb);
2748         }
2749 }
2750
2751 void set_page_extent_mapped(struct page *page)
2752 {
2753         if (!PagePrivate(page)) {
2754                 SetPagePrivate(page);
2755                 page_cache_get(page);
2756                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2757         }
2758 }
2759
2760 static struct extent_map *
2761 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2762                  u64 start, u64 len, get_extent_t *get_extent,
2763                  struct extent_map **em_cached)
2764 {
2765         struct extent_map *em;
2766
2767         if (em_cached && *em_cached) {
2768                 em = *em_cached;
2769                 if (extent_map_in_tree(em) && start >= em->start &&
2770                     start < extent_map_end(em)) {
2771                         atomic_inc(&em->refs);
2772                         return em;
2773                 }
2774
2775                 free_extent_map(em);
2776                 *em_cached = NULL;
2777         }
2778
2779         em = get_extent(inode, page, pg_offset, start, len, 0);
2780         if (em_cached && !IS_ERR_OR_NULL(em)) {
2781                 BUG_ON(*em_cached);
2782                 atomic_inc(&em->refs);
2783                 *em_cached = em;
2784         }
2785         return em;
2786 }
2787 /*
2788  * basic readpage implementation.  Locked extent state structs are inserted
2789  * into the tree that are removed when the IO is done (by the end_io
2790  * handlers)
2791  * XXX JDM: This needs looking at to ensure proper page locking
2792  */
2793 static int __do_readpage(struct extent_io_tree *tree,
2794                          struct page *page,
2795                          get_extent_t *get_extent,
2796                          struct extent_map **em_cached,
2797                          struct bio **bio, int mirror_num,
2798                          unsigned long *bio_flags, int rw)
2799 {
2800         struct inode *inode = page->mapping->host;
2801         u64 start = page_offset(page);
2802         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2803         u64 end;
2804         u64 cur = start;
2805         u64 extent_offset;
2806         u64 last_byte = i_size_read(inode);
2807         u64 block_start;
2808         u64 cur_end;
2809         sector_t sector;
2810         struct extent_map *em;
2811         struct block_device *bdev;
2812         int ret;
2813         int nr = 0;
2814         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2815         size_t pg_offset = 0;
2816         size_t iosize;
2817         size_t disk_io_size;
2818         size_t blocksize = inode->i_sb->s_blocksize;
2819         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2820
2821         set_page_extent_mapped(page);
2822
2823         end = page_end;
2824         if (!PageUptodate(page)) {
2825                 if (cleancache_get_page(page) == 0) {
2826                         BUG_ON(blocksize != PAGE_SIZE);
2827                         unlock_extent(tree, start, end);
2828                         goto out;
2829                 }
2830         }
2831
2832         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2833                 char *userpage;
2834                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2835
2836                 if (zero_offset) {
2837                         iosize = PAGE_CACHE_SIZE - zero_offset;
2838                         userpage = kmap_atomic(page);
2839                         memset(userpage + zero_offset, 0, iosize);
2840                         flush_dcache_page(page);
2841                         kunmap_atomic(userpage);
2842                 }
2843         }
2844         while (cur <= end) {
2845                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2846
2847                 if (cur >= last_byte) {
2848                         char *userpage;
2849                         struct extent_state *cached = NULL;
2850
2851                         iosize = PAGE_CACHE_SIZE - pg_offset;
2852                         userpage = kmap_atomic(page);
2853                         memset(userpage + pg_offset, 0, iosize);
2854                         flush_dcache_page(page);
2855                         kunmap_atomic(userpage);
2856                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2857                                             &cached, GFP_NOFS);
2858                         if (!parent_locked)
2859                                 unlock_extent_cached(tree, cur,
2860                                                      cur + iosize - 1,
2861                                                      &cached, GFP_NOFS);
2862                         break;
2863                 }
2864                 em = __get_extent_map(inode, page, pg_offset, cur,
2865                                       end - cur + 1, get_extent, em_cached);
2866                 if (IS_ERR_OR_NULL(em)) {
2867                         SetPageError(page);
2868                         if (!parent_locked)
2869                                 unlock_extent(tree, cur, end);
2870                         break;
2871                 }
2872                 extent_offset = cur - em->start;
2873                 BUG_ON(extent_map_end(em) <= cur);
2874                 BUG_ON(end < cur);
2875
2876                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2877                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2878                         extent_set_compress_type(&this_bio_flag,
2879                                                  em->compress_type);
2880                 }
2881
2882                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2883                 cur_end = min(extent_map_end(em) - 1, end);
2884                 iosize = ALIGN(iosize, blocksize);
2885                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2886                         disk_io_size = em->block_len;
2887                         sector = em->block_start >> 9;
2888                 } else {
2889                         sector = (em->block_start + extent_offset) >> 9;
2890                         disk_io_size = iosize;
2891                 }
2892                 bdev = em->bdev;
2893                 block_start = em->block_start;
2894                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2895                         block_start = EXTENT_MAP_HOLE;
2896                 free_extent_map(em);
2897                 em = NULL;
2898
2899                 /* we've found a hole, just zero and go on */
2900                 if (block_start == EXTENT_MAP_HOLE) {
2901                         char *userpage;
2902                         struct extent_state *cached = NULL;
2903
2904                         userpage = kmap_atomic(page);
2905                         memset(userpage + pg_offset, 0, iosize);
2906                         flush_dcache_page(page);
2907                         kunmap_atomic(userpage);
2908
2909                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2910                                             &cached, GFP_NOFS);
2911                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2912                                              &cached, GFP_NOFS);
2913                         cur = cur + iosize;
2914                         pg_offset += iosize;
2915                         continue;
2916                 }
2917                 /* the get_extent function already copied into the page */
2918                 if (test_range_bit(tree, cur, cur_end,
2919                                    EXTENT_UPTODATE, 1, NULL)) {
2920                         check_page_uptodate(tree, page);
2921                         if (!parent_locked)
2922                                 unlock_extent(tree, cur, cur + iosize - 1);
2923                         cur = cur + iosize;
2924                         pg_offset += iosize;
2925                         continue;
2926                 }
2927                 /* we have an inline extent but it didn't get marked up
2928                  * to date.  Error out
2929                  */
2930                 if (block_start == EXTENT_MAP_INLINE) {
2931                         SetPageError(page);
2932                         if (!parent_locked)
2933                                 unlock_extent(tree, cur, cur + iosize - 1);
2934                         cur = cur + iosize;
2935                         pg_offset += iosize;
2936                         continue;
2937                 }
2938
2939                 pnr -= page->index;
2940                 ret = submit_extent_page(rw, tree, page,
2941                                          sector, disk_io_size, pg_offset,
2942                                          bdev, bio, pnr,
2943                                          end_bio_extent_readpage, mirror_num,
2944                                          *bio_flags,
2945                                          this_bio_flag);
2946                 if (!ret) {
2947                         nr++;
2948                         *bio_flags = this_bio_flag;
2949                 } else {
2950                         SetPageError(page);
2951                         if (!parent_locked)
2952                                 unlock_extent(tree, cur, cur + iosize - 1);
2953                 }
2954                 cur = cur + iosize;
2955                 pg_offset += iosize;
2956         }
2957 out:
2958         if (!nr) {
2959                 if (!PageError(page))
2960                         SetPageUptodate(page);
2961                 unlock_page(page);
2962         }
2963         return 0;
2964 }
2965
2966 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
2967                                              struct page *pages[], int nr_pages,
2968                                              u64 start, u64 end,
2969                                              get_extent_t *get_extent,
2970                                              struct extent_map **em_cached,
2971                                              struct bio **bio, int mirror_num,
2972                                              unsigned long *bio_flags, int rw)
2973 {
2974         struct inode *inode;
2975         struct btrfs_ordered_extent *ordered;
2976         int index;
2977
2978         inode = pages[0]->mapping->host;
2979         while (1) {
2980                 lock_extent(tree, start, end);
2981                 ordered = btrfs_lookup_ordered_range(inode, start,
2982                                                      end - start + 1);
2983                 if (!ordered)
2984                         break;
2985                 unlock_extent(tree, start, end);
2986                 btrfs_start_ordered_extent(inode, ordered, 1);
2987                 btrfs_put_ordered_extent(ordered);
2988         }
2989
2990         for (index = 0; index < nr_pages; index++) {
2991                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
2992                               mirror_num, bio_flags, rw);
2993                 page_cache_release(pages[index]);
2994         }
2995 }
2996
2997 static void __extent_readpages(struct extent_io_tree *tree,
2998                                struct page *pages[],
2999                                int nr_pages, get_extent_t *get_extent,
3000                                struct extent_map **em_cached,
3001                                struct bio **bio, int mirror_num,
3002                                unsigned long *bio_flags, int rw)
3003 {
3004         u64 start = 0;
3005         u64 end = 0;
3006         u64 page_start;
3007         int index;
3008         int first_index = 0;
3009
3010         for (index = 0; index < nr_pages; index++) {
3011                 page_start = page_offset(pages[index]);
3012                 if (!end) {
3013                         start = page_start;
3014                         end = start + PAGE_CACHE_SIZE - 1;
3015                         first_index = index;
3016                 } else if (end + 1 == page_start) {
3017                         end += PAGE_CACHE_SIZE;
3018                 } else {
3019                         __do_contiguous_readpages(tree, &pages[first_index],
3020                                                   index - first_index, start,
3021                                                   end, get_extent, em_cached,
3022                                                   bio, mirror_num, bio_flags,
3023                                                   rw);
3024                         start = page_start;
3025                         end = start + PAGE_CACHE_SIZE - 1;
3026                         first_index = index;
3027                 }
3028         }
3029
3030         if (end)
3031                 __do_contiguous_readpages(tree, &pages[first_index],
3032                                           index - first_index, start,
3033                                           end, get_extent, em_cached, bio,
3034                                           mirror_num, bio_flags, rw);
3035 }
3036
3037 static int __extent_read_full_page(struct extent_io_tree *tree,
3038                                    struct page *page,
3039                                    get_extent_t *get_extent,
3040                                    struct bio **bio, int mirror_num,
3041                                    unsigned long *bio_flags, int rw)
3042 {
3043         struct inode *inode = page->mapping->host;
3044         struct btrfs_ordered_extent *ordered;
3045         u64 start = page_offset(page);
3046         u64 end = start + PAGE_CACHE_SIZE - 1;
3047         int ret;
3048
3049         while (1) {
3050                 lock_extent(tree, start, end);
3051                 ordered = btrfs_lookup_ordered_extent(inode, start);
3052                 if (!ordered)
3053                         break;
3054                 unlock_extent(tree, start, end);
3055                 btrfs_start_ordered_extent(inode, ordered, 1);
3056                 btrfs_put_ordered_extent(ordered);
3057         }
3058
3059         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3060                             bio_flags, rw);
3061         return ret;
3062 }
3063
3064 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3065                             get_extent_t *get_extent, int mirror_num)
3066 {
3067         struct bio *bio = NULL;
3068         unsigned long bio_flags = 0;
3069         int ret;
3070
3071         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3072                                       &bio_flags, READ);
3073         if (bio)
3074                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3075         return ret;
3076 }
3077
3078 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3079                                  get_extent_t *get_extent, int mirror_num)
3080 {
3081         struct bio *bio = NULL;
3082         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3083         int ret;
3084
3085         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3086                                       &bio_flags, READ);
3087         if (bio)
3088                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3089         return ret;
3090 }
3091
3092 static noinline void update_nr_written(struct page *page,
3093                                       struct writeback_control *wbc,
3094                                       unsigned long nr_written)
3095 {
3096         wbc->nr_to_write -= nr_written;
3097         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3098             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3099                 page->mapping->writeback_index = page->index + nr_written;
3100 }
3101
3102 /*
3103  * the writepage semantics are similar to regular writepage.  extent
3104  * records are inserted to lock ranges in the tree, and as dirty areas
3105  * are found, they are marked writeback.  Then the lock bits are removed
3106  * and the end_io handler clears the writeback ranges
3107  */
3108 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3109                               void *data)
3110 {
3111         struct inode *inode = page->mapping->host;
3112         struct extent_page_data *epd = data;
3113         struct extent_io_tree *tree = epd->tree;
3114         u64 start = page_offset(page);
3115         u64 delalloc_start;
3116         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3117         u64 end;
3118         u64 cur = start;
3119         u64 extent_offset;
3120         u64 last_byte = i_size_read(inode);
3121         u64 block_start;
3122         u64 iosize;
3123         sector_t sector;
3124         struct extent_state *cached_state = NULL;
3125         struct extent_map *em;
3126         struct block_device *bdev;
3127         int ret;
3128         int nr = 0;
3129         size_t pg_offset = 0;
3130         size_t blocksize;
3131         loff_t i_size = i_size_read(inode);
3132         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3133         u64 nr_delalloc;
3134         u64 delalloc_end;
3135         int page_started;
3136         int compressed;
3137         int write_flags;
3138         unsigned long nr_written = 0;
3139         bool fill_delalloc = true;
3140
3141         if (wbc->sync_mode == WB_SYNC_ALL)
3142                 write_flags = WRITE_SYNC;
3143         else
3144                 write_flags = WRITE;
3145
3146         trace___extent_writepage(page, inode, wbc);
3147
3148         WARN_ON(!PageLocked(page));
3149
3150         ClearPageError(page);
3151
3152         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3153         if (page->index > end_index ||
3154            (page->index == end_index && !pg_offset)) {
3155                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3156                 unlock_page(page);
3157                 return 0;
3158         }
3159
3160         if (page->index == end_index) {
3161                 char *userpage;
3162
3163                 userpage = kmap_atomic(page);
3164                 memset(userpage + pg_offset, 0,
3165                        PAGE_CACHE_SIZE - pg_offset);
3166                 kunmap_atomic(userpage);
3167                 flush_dcache_page(page);
3168         }
3169         pg_offset = 0;
3170
3171         set_page_extent_mapped(page);
3172
3173         if (!tree->ops || !tree->ops->fill_delalloc)
3174                 fill_delalloc = false;
3175
3176         delalloc_start = start;
3177         delalloc_end = 0;
3178         page_started = 0;
3179         if (!epd->extent_locked && fill_delalloc) {
3180                 u64 delalloc_to_write = 0;
3181                 /*
3182                  * make sure the wbc mapping index is at least updated
3183                  * to this page.
3184                  */
3185                 update_nr_written(page, wbc, 0);
3186
3187                 while (delalloc_end < page_end) {
3188                         nr_delalloc = find_lock_delalloc_range(inode, tree,
3189                                                        page,
3190                                                        &delalloc_start,
3191                                                        &delalloc_end,
3192                                                        128 * 1024 * 1024);
3193                         if (nr_delalloc == 0) {
3194                                 delalloc_start = delalloc_end + 1;
3195                                 continue;
3196                         }
3197                         ret = tree->ops->fill_delalloc(inode, page,
3198                                                        delalloc_start,
3199                                                        delalloc_end,
3200                                                        &page_started,
3201                                                        &nr_written);
3202                         /* File system has been set read-only */
3203                         if (ret) {
3204                                 SetPageError(page);
3205                                 goto done;
3206                         }
3207                         /*
3208                          * delalloc_end is already one less than the total
3209                          * length, so we don't subtract one from
3210                          * PAGE_CACHE_SIZE
3211                          */
3212                         delalloc_to_write += (delalloc_end - delalloc_start +
3213                                               PAGE_CACHE_SIZE) >>
3214                                               PAGE_CACHE_SHIFT;
3215                         delalloc_start = delalloc_end + 1;
3216                 }
3217                 if (wbc->nr_to_write < delalloc_to_write) {
3218                         int thresh = 8192;
3219
3220                         if (delalloc_to_write < thresh * 2)
3221                                 thresh = delalloc_to_write;
3222                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
3223                                                  thresh);
3224                 }
3225
3226                 /* did the fill delalloc function already unlock and start
3227                  * the IO?
3228                  */
3229                 if (page_started) {
3230                         ret = 0;
3231                         /*
3232                          * we've unlocked the page, so we can't update
3233                          * the mapping's writeback index, just update
3234                          * nr_to_write.
3235                          */
3236                         wbc->nr_to_write -= nr_written;
3237                         goto done_unlocked;
3238                 }
3239         }
3240         if (tree->ops && tree->ops->writepage_start_hook) {
3241                 ret = tree->ops->writepage_start_hook(page, start,
3242                                                       page_end);
3243                 if (ret) {
3244                         /* Fixup worker will requeue */
3245                         if (ret == -EBUSY)
3246                                 wbc->pages_skipped++;
3247                         else
3248                                 redirty_page_for_writepage(wbc, page);
3249                         update_nr_written(page, wbc, nr_written);
3250                         unlock_page(page);
3251                         ret = 0;
3252                         goto done_unlocked;
3253                 }
3254         }
3255
3256         /*
3257          * we don't want to touch the inode after unlocking the page,
3258          * so we update the mapping writeback index now
3259          */
3260         update_nr_written(page, wbc, nr_written + 1);
3261
3262         end = page_end;
3263         if (last_byte <= start) {
3264                 if (tree->ops && tree->ops->writepage_end_io_hook)
3265                         tree->ops->writepage_end_io_hook(page, start,
3266                                                          page_end, NULL, 1);
3267                 goto done;
3268         }
3269
3270         blocksize = inode->i_sb->s_blocksize;
3271
3272         while (cur <= end) {
3273                 if (cur >= last_byte) {
3274                         if (tree->ops && tree->ops->writepage_end_io_hook)
3275                                 tree->ops->writepage_end_io_hook(page, cur,
3276                                                          page_end, NULL, 1);
3277                         break;
3278                 }
3279                 em = epd->get_extent(inode, page, pg_offset, cur,
3280                                      end - cur + 1, 1);
3281                 if (IS_ERR_OR_NULL(em)) {
3282                         SetPageError(page);
3283                         ret = PTR_ERR_OR_ZERO(em);
3284                         break;
3285                 }
3286
3287                 extent_offset = cur - em->start;
3288                 BUG_ON(extent_map_end(em) <= cur);
3289                 BUG_ON(end < cur);
3290                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
3291                 iosize = ALIGN(iosize, blocksize);
3292                 sector = (em->block_start + extent_offset) >> 9;
3293                 bdev = em->bdev;
3294                 block_start = em->block_start;
3295                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3296                 free_extent_map(em);
3297                 em = NULL;
3298
3299                 /*
3300                  * compressed and inline extents are written through other
3301                  * paths in the FS
3302                  */
3303                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3304                     block_start == EXTENT_MAP_INLINE) {
3305                         /*
3306                          * end_io notification does not happen here for
3307                          * compressed extents
3308                          */
3309                         if (!compressed && tree->ops &&
3310                             tree->ops->writepage_end_io_hook)
3311                                 tree->ops->writepage_end_io_hook(page, cur,
3312                                                          cur + iosize - 1,
3313                                                          NULL, 1);
3314                         else if (compressed) {
3315                                 /* we don't want to end_page_writeback on
3316                                  * a compressed extent.  this happens
3317                                  * elsewhere
3318                                  */
3319                                 nr++;
3320                         }
3321
3322                         cur += iosize;
3323                         pg_offset += iosize;
3324                         continue;
3325                 }
3326                 /* leave this out until we have a page_mkwrite call */
3327                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
3328                                    EXTENT_DIRTY, 0, NULL)) {
3329                         cur = cur + iosize;
3330                         pg_offset += iosize;
3331                         continue;
3332                 }
3333
3334                 if (tree->ops && tree->ops->writepage_io_hook) {
3335                         ret = tree->ops->writepage_io_hook(page, cur,
3336                                                 cur + iosize - 1);
3337                 } else {
3338                         ret = 0;
3339                 }
3340                 if (ret) {
3341                         SetPageError(page);
3342                 } else {
3343                         unsigned long max_nr = end_index + 1;
3344
3345                         set_range_writeback(tree, cur, cur + iosize - 1);
3346                         if (!PageWriteback(page)) {
3347                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
3348                                            "page %lu not writeback, cur %llu end %llu",
3349                                        page->index, cur, end);
3350                         }
3351
3352                         ret = submit_extent_page(write_flags, tree, page,
3353                                                  sector, iosize, pg_offset,
3354                                                  bdev, &epd->bio, max_nr,
3355                                                  end_bio_extent_writepage,
3356                                                  0, 0, 0);
3357                         if (ret)
3358                                 SetPageError(page);
3359                 }
3360                 cur = cur + iosize;
3361                 pg_offset += iosize;
3362                 nr++;
3363         }
3364 done:
3365         if (nr == 0) {
3366                 /* make sure the mapping tag for page dirty gets cleared */
3367                 set_page_writeback(page);
3368                 end_page_writeback(page);
3369         }
3370         if (PageError(page)) {
3371                 ret = ret < 0 ? ret : -EIO;
3372                 end_extent_writepage(page, ret, start, page_end);
3373         }
3374         unlock_page(page);
3375
3376 done_unlocked:
3377
3378         /* drop our reference on any cached states */
3379         free_extent_state(cached_state);
3380         return ret;
3381 }
3382
3383 static int eb_wait(void *word)
3384 {
3385         io_schedule();
3386         return 0;
3387 }
3388
3389 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3390 {
3391         wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
3392                     TASK_UNINTERRUPTIBLE);
3393 }
3394
3395 static int lock_extent_buffer_for_io(struct extent_buffer *eb,
3396                                      struct btrfs_fs_info *fs_info,
3397                                      struct extent_page_data *epd)
3398 {
3399         unsigned long i, num_pages;
3400         int flush = 0;
3401         int ret = 0;
3402
3403         if (!btrfs_try_tree_write_lock(eb)) {
3404                 flush = 1;
3405                 flush_write_bio(epd);
3406                 btrfs_tree_lock(eb);
3407         }
3408
3409         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3410                 btrfs_tree_unlock(eb);
3411                 if (!epd->sync_io)
3412                         return 0;
3413                 if (!flush) {
3414                         flush_write_bio(epd);
3415                         flush = 1;
3416                 }
3417                 while (1) {
3418                         wait_on_extent_buffer_writeback(eb);
3419                         btrfs_tree_lock(eb);
3420                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3421                                 break;
3422                         btrfs_tree_unlock(eb);
3423                 }
3424         }
3425
3426         /*
3427          * We need to do this to prevent races in people who check if the eb is
3428          * under IO since we can end up having no IO bits set for a short period
3429          * of time.
3430          */
3431         spin_lock(&eb->refs_lock);
3432         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3433                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3434                 spin_unlock(&eb->refs_lock);
3435                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3436                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3437                                      -eb->len,
3438                                      fs_info->dirty_metadata_batch);
3439                 ret = 1;
3440         } else {
3441                 spin_unlock(&eb->refs_lock);
3442         }
3443
3444         btrfs_tree_unlock(eb);
3445
3446         if (!ret)
3447                 return ret;
3448
3449         num_pages = num_extent_pages(eb->start, eb->len);
3450         for (i = 0; i < num_pages; i++) {
3451                 struct page *p = extent_buffer_page(eb, i);
3452
3453                 if (!trylock_page(p)) {
3454                         if (!flush) {
3455                                 flush_write_bio(epd);
3456                                 flush = 1;
3457                         }
3458                         lock_page(p);
3459                 }
3460         }
3461
3462         return ret;
3463 }
3464
3465 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3466 {
3467         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3468         smp_mb__after_clear_bit();
3469         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3470 }
3471
3472 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3473 {
3474         struct bio_vec *bvec;
3475         struct extent_buffer *eb;
3476         int i, done;
3477
3478         bio_for_each_segment_all(bvec, bio, i) {
3479                 struct page *page = bvec->bv_page;
3480
3481                 eb = (struct extent_buffer *)page->private;
3482                 BUG_ON(!eb);
3483                 done = atomic_dec_and_test(&eb->io_pages);
3484
3485                 if (err || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3486                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3487                         ClearPageUptodate(page);
3488                         SetPageError(page);
3489                 }
3490
3491                 end_page_writeback(page);
3492
3493                 if (!done)
3494                         continue;
3495
3496                 end_extent_buffer_writeback(eb);
3497         }
3498
3499         bio_put(bio);
3500 }
3501
3502 static int write_one_eb(struct extent_buffer *eb,
3503                         struct btrfs_fs_info *fs_info,
3504                         struct writeback_control *wbc,
3505                         struct extent_page_data *epd)
3506 {
3507         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3508         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3509         u64 offset = eb->start;
3510         unsigned long i, num_pages;
3511         unsigned long bio_flags = 0;
3512         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3513         int ret = 0;
3514
3515         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3516         num_pages = num_extent_pages(eb->start, eb->len);
3517         atomic_set(&eb->io_pages, num_pages);
3518         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3519                 bio_flags = EXTENT_BIO_TREE_LOG;
3520
3521         for (i = 0; i < num_pages; i++) {
3522                 struct page *p = extent_buffer_page(eb, i);
3523
3524                 clear_page_dirty_for_io(p);
3525                 set_page_writeback(p);
3526                 ret = submit_extent_page(rw, tree, p, offset >> 9,
3527                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3528                                          -1, end_bio_extent_buffer_writepage,
3529                                          0, epd->bio_flags, bio_flags);
3530                 epd->bio_flags = bio_flags;
3531                 if (ret) {
3532                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3533                         SetPageError(p);
3534                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3535                                 end_extent_buffer_writeback(eb);
3536                         ret = -EIO;
3537                         break;
3538                 }
3539                 offset += PAGE_CACHE_SIZE;
3540                 update_nr_written(p, wbc, 1);
3541                 unlock_page(p);
3542         }
3543
3544         if (unlikely(ret)) {
3545                 for (; i < num_pages; i++) {
3546                         struct page *p = extent_buffer_page(eb, i);
3547                         unlock_page(p);
3548                 }
3549         }
3550
3551         return ret;
3552 }
3553
3554 int btree_write_cache_pages(struct address_space *mapping,
3555                                    struct writeback_control *wbc)
3556 {
3557         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3558         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3559         struct extent_buffer *eb, *prev_eb = NULL;
3560         struct extent_page_data epd = {
3561                 .bio = NULL,
3562                 .tree = tree,
3563                 .extent_locked = 0,
3564                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3565                 .bio_flags = 0,
3566         };
3567         int ret = 0;
3568         int done = 0;
3569         int nr_to_write_done = 0;
3570         struct pagevec pvec;
3571         int nr_pages;
3572         pgoff_t index;
3573         pgoff_t end;            /* Inclusive */
3574         int scanned = 0;
3575         int tag;
3576
3577         pagevec_init(&pvec, 0);
3578         if (wbc->range_cyclic) {
3579                 index = mapping->writeback_index; /* Start from prev offset */
3580                 end = -1;
3581         } else {
3582                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3583                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3584                 scanned = 1;
3585         }
3586         if (wbc->sync_mode == WB_SYNC_ALL)
3587                 tag = PAGECACHE_TAG_TOWRITE;
3588         else
3589                 tag = PAGECACHE_TAG_DIRTY;
3590 retry:
3591         if (wbc->sync_mode == WB_SYNC_ALL)
3592                 tag_pages_for_writeback(mapping, index, end);
3593         while (!done && !nr_to_write_done && (index <= end) &&
3594                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3595                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3596                 unsigned i;
3597
3598                 scanned = 1;
3599                 for (i = 0; i < nr_pages; i++) {
3600                         struct page *page = pvec.pages[i];
3601
3602                         if (!PagePrivate(page))
3603                                 continue;
3604
3605                         if (!wbc->range_cyclic && page->index > end) {
3606                                 done = 1;
3607                                 break;
3608                         }
3609
3610                         spin_lock(&mapping->private_lock);
3611                         if (!PagePrivate(page)) {
3612                                 spin_unlock(&mapping->private_lock);
3613                                 continue;
3614                         }
3615
3616                         eb = (struct extent_buffer *)page->private;
3617
3618                         /*
3619                          * Shouldn't happen and normally this would be a BUG_ON
3620                          * but no sense in crashing the users box for something
3621                          * we can survive anyway.
3622                          */
3623                         if (WARN_ON(!eb)) {
3624                                 spin_unlock(&mapping->private_lock);
3625                                 continue;
3626                         }
3627
3628                         if (eb == prev_eb) {
3629                                 spin_unlock(&mapping->private_lock);
3630                                 continue;
3631                         }
3632
3633                         ret = atomic_inc_not_zero(&eb->refs);
3634                         spin_unlock(&mapping->private_lock);
3635                         if (!ret)
3636                                 continue;
3637
3638                         prev_eb = eb;
3639                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3640                         if (!ret) {
3641                                 free_extent_buffer(eb);
3642                                 continue;
3643                         }
3644
3645                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3646                         if (ret) {
3647                                 done = 1;
3648                                 free_extent_buffer(eb);
3649                                 break;
3650                         }
3651                         free_extent_buffer(eb);
3652
3653                         /*
3654                          * the filesystem may choose to bump up nr_to_write.
3655                          * We have to make sure to honor the new nr_to_write
3656                          * at any time
3657                          */
3658                         nr_to_write_done = wbc->nr_to_write <= 0;
3659                 }
3660                 pagevec_release(&pvec);
3661                 cond_resched();
3662         }
3663         if (!scanned && !done) {
3664                 /*
3665                  * We hit the last page and there is more work to be done: wrap
3666                  * back to the start of the file
3667                  */
3668                 scanned = 1;
3669                 index = 0;
3670                 goto retry;
3671         }
3672         flush_write_bio(&epd);
3673         return ret;
3674 }
3675
3676 /**
3677  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3678  * @mapping: address space structure to write
3679  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3680  * @writepage: function called for each page
3681  * @data: data passed to writepage function
3682  *
3683  * If a page is already under I/O, write_cache_pages() skips it, even
3684  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3685  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3686  * and msync() need to guarantee that all the data which was dirty at the time
3687  * the call was made get new I/O started against them.  If wbc->sync_mode is
3688  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3689  * existing IO to complete.
3690  */
3691 static int extent_write_cache_pages(struct extent_io_tree *tree,
3692                              struct address_space *mapping,
3693                              struct writeback_control *wbc,
3694                              writepage_t writepage, void *data,
3695                              void (*flush_fn)(void *))
3696 {
3697         struct inode *inode = mapping->host;
3698         int ret = 0;
3699         int done = 0;
3700         int err = 0;
3701         int nr_to_write_done = 0;
3702         struct pagevec pvec;
3703         int nr_pages;
3704         pgoff_t index;
3705         pgoff_t end;            /* Inclusive */
3706         int scanned = 0;
3707         int tag;
3708
3709         /*
3710          * We have to hold onto the inode so that ordered extents can do their
3711          * work when the IO finishes.  The alternative to this is failing to add
3712          * an ordered extent if the igrab() fails there and that is a huge pain
3713          * to deal with, so instead just hold onto the inode throughout the
3714          * writepages operation.  If it fails here we are freeing up the inode
3715          * anyway and we'd rather not waste our time writing out stuff that is
3716          * going to be truncated anyway.
3717          */
3718         if (!igrab(inode))
3719                 return 0;
3720
3721         pagevec_init(&pvec, 0);
3722         if (wbc->range_cyclic) {
3723                 index = mapping->writeback_index; /* Start from prev offset */
3724                 end = -1;
3725         } else {
3726                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3727                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3728                 scanned = 1;
3729         }
3730         if (wbc->sync_mode == WB_SYNC_ALL)
3731                 tag = PAGECACHE_TAG_TOWRITE;
3732         else
3733                 tag = PAGECACHE_TAG_DIRTY;
3734 retry:
3735         if (wbc->sync_mode == WB_SYNC_ALL)
3736                 tag_pages_for_writeback(mapping, index, end);
3737         while (!done && !nr_to_write_done && (index <= end) &&
3738                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3739                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3740                 unsigned i;
3741
3742                 scanned = 1;
3743                 for (i = 0; i < nr_pages; i++) {
3744                         struct page *page = pvec.pages[i];
3745
3746                         /*
3747                          * At this point we hold neither mapping->tree_lock nor
3748                          * lock on the page itself: the page may be truncated or
3749                          * invalidated (changing page->mapping to NULL), or even
3750                          * swizzled back from swapper_space to tmpfs file
3751                          * mapping
3752                          */
3753                         if (!trylock_page(page)) {
3754                                 flush_fn(data);
3755                                 lock_page(page);
3756                         }
3757
3758                         if (unlikely(page->mapping != mapping)) {
3759                                 unlock_page(page);
3760                                 continue;
3761                         }
3762
3763                         if (!wbc->range_cyclic && page->index > end) {
3764                                 done = 1;
3765                                 unlock_page(page);
3766                                 continue;
3767                         }
3768
3769                         if (wbc->sync_mode != WB_SYNC_NONE) {
3770                                 if (PageWriteback(page))
3771                                         flush_fn(data);
3772                                 wait_on_page_writeback(page);
3773                         }
3774
3775                         if (PageWriteback(page) ||
3776                             !clear_page_dirty_for_io(page)) {
3777                                 unlock_page(page);
3778                                 continue;
3779                         }
3780
3781                         ret = (*writepage)(page, wbc, data);
3782
3783                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3784                                 unlock_page(page);
3785                                 ret = 0;
3786                         }
3787                         if (!err && ret < 0)
3788                                 err = ret;
3789
3790                         /*
3791                          * the filesystem may choose to bump up nr_to_write.
3792                          * We have to make sure to honor the new nr_to_write
3793                          * at any time
3794                          */
3795                         nr_to_write_done = wbc->nr_to_write <= 0;
3796                 }
3797                 pagevec_release(&pvec);
3798                 cond_resched();
3799         }
3800         if (!scanned && !done && !err) {
3801                 /*
3802                  * We hit the last page and there is more work to be done: wrap
3803                  * back to the start of the file
3804                  */
3805                 scanned = 1;
3806                 index = 0;
3807                 goto retry;
3808         }
3809         btrfs_add_delayed_iput(inode);
3810         return err;
3811 }
3812
3813 static void flush_epd_write_bio(struct extent_page_data *epd)
3814 {
3815         if (epd->bio) {
3816                 int rw = WRITE;
3817                 int ret;
3818
3819                 if (epd->sync_io)
3820                         rw = WRITE_SYNC;
3821
3822                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
3823                 BUG_ON(ret < 0); /* -ENOMEM */
3824                 epd->bio = NULL;
3825         }
3826 }
3827
3828 static noinline void flush_write_bio(void *data)
3829 {
3830         struct extent_page_data *epd = data;
3831         flush_epd_write_bio(epd);
3832 }
3833
3834 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3835                           get_extent_t *get_extent,
3836                           struct writeback_control *wbc)
3837 {
3838         int ret;
3839         struct extent_page_data epd = {
3840                 .bio = NULL,
3841                 .tree = tree,
3842                 .get_extent = get_extent,
3843                 .extent_locked = 0,
3844                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3845                 .bio_flags = 0,
3846         };
3847
3848         ret = __extent_writepage(page, wbc, &epd);
3849
3850         flush_epd_write_bio(&epd);
3851         return ret;
3852 }
3853
3854 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3855                               u64 start, u64 end, get_extent_t *get_extent,
3856                               int mode)
3857 {
3858         int ret = 0;
3859         struct address_space *mapping = inode->i_mapping;
3860         struct page *page;
3861         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3862                 PAGE_CACHE_SHIFT;
3863
3864         struct extent_page_data epd = {
3865                 .bio = NULL,
3866                 .tree = tree,
3867                 .get_extent = get_extent,
3868                 .extent_locked = 1,
3869                 .sync_io = mode == WB_SYNC_ALL,
3870                 .bio_flags = 0,
3871         };
3872         struct writeback_control wbc_writepages = {
3873                 .sync_mode      = mode,
3874                 .nr_to_write    = nr_pages * 2,
3875                 .range_start    = start,
3876                 .range_end      = end + 1,
3877         };
3878
3879         while (start <= end) {
3880                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3881                 if (clear_page_dirty_for_io(page))
3882                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3883                 else {
3884                         if (tree->ops && tree->ops->writepage_end_io_hook)
3885                                 tree->ops->writepage_end_io_hook(page, start,
3886                                                  start + PAGE_CACHE_SIZE - 1,
3887                                                  NULL, 1);
3888                         unlock_page(page);
3889                 }
3890                 page_cache_release(page);
3891                 start += PAGE_CACHE_SIZE;
3892         }
3893
3894         flush_epd_write_bio(&epd);
3895         return ret;
3896 }
3897
3898 int extent_writepages(struct extent_io_tree *tree,
3899                       struct address_space *mapping,
3900                       get_extent_t *get_extent,
3901                       struct writeback_control *wbc)
3902 {
3903         int ret = 0;
3904         struct extent_page_data epd = {
3905                 .bio = NULL,
3906                 .tree = tree,
3907                 .get_extent = get_extent,
3908                 .extent_locked = 0,
3909                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3910                 .bio_flags = 0,
3911         };
3912
3913         ret = extent_write_cache_pages(tree, mapping, wbc,
3914                                        __extent_writepage, &epd,
3915                                        flush_write_bio);
3916         flush_epd_write_bio(&epd);
3917         return ret;
3918 }
3919
3920 int extent_readpages(struct extent_io_tree *tree,
3921                      struct address_space *mapping,
3922                      struct list_head *pages, unsigned nr_pages,
3923                      get_extent_t get_extent)
3924 {
3925         struct bio *bio = NULL;
3926         unsigned page_idx;
3927         unsigned long bio_flags = 0;
3928         struct page *pagepool[16];
3929         struct page *page;
3930         struct extent_map *em_cached = NULL;
3931         int nr = 0;
3932
3933         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3934                 page = list_entry(pages->prev, struct page, lru);
3935
3936                 prefetchw(&page->flags);
3937                 list_del(&page->lru);
3938                 if (add_to_page_cache_lru(page, mapping,
3939                                         page->index, GFP_NOFS)) {
3940                         page_cache_release(page);
3941                         continue;
3942                 }
3943
3944                 pagepool[nr++] = page;
3945                 if (nr < ARRAY_SIZE(pagepool))
3946                         continue;
3947                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
3948                                    &bio, 0, &bio_flags, READ);
3949                 nr = 0;
3950         }
3951         if (nr)
3952                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
3953                                    &bio, 0, &bio_flags, READ);
3954
3955         if (em_cached)
3956                 free_extent_map(em_cached);
3957
3958         BUG_ON(!list_empty(pages));
3959         if (bio)
3960                 return submit_one_bio(READ, bio, 0, bio_flags);
3961         return 0;
3962 }
3963
3964 /*
3965  * basic invalidatepage code, this waits on any locked or writeback
3966  * ranges corresponding to the page, and then deletes any extent state
3967  * records from the tree
3968  */
3969 int extent_invalidatepage(struct extent_io_tree *tree,
3970                           struct page *page, unsigned long offset)
3971 {
3972         struct extent_state *cached_state = NULL;
3973         u64 start = page_offset(page);
3974         u64 end = start + PAGE_CACHE_SIZE - 1;
3975         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3976
3977         start += ALIGN(offset, blocksize);
3978         if (start > end)
3979                 return 0;
3980
3981         lock_extent_bits(tree, start, end, 0, &cached_state);
3982         wait_on_page_writeback(page);
3983         clear_extent_bit(tree, start, end,
3984                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3985                          EXTENT_DO_ACCOUNTING,
3986                          1, 1, &cached_state, GFP_NOFS);
3987         return 0;
3988 }
3989
3990 /*
3991  * a helper for releasepage, this tests for areas of the page that
3992  * are locked or under IO and drops the related state bits if it is safe
3993  * to drop the page.
3994  */
3995 static int try_release_extent_state(struct extent_map_tree *map,
3996                                     struct extent_io_tree *tree,
3997                                     struct page *page, gfp_t mask)
3998 {
3999         u64 start = page_offset(page);
4000         u64 end = start + PAGE_CACHE_SIZE - 1;
4001         int ret = 1;
4002
4003         if (test_range_bit(tree, start, end,
4004                            EXTENT_IOBITS, 0, NULL))
4005                 ret = 0;
4006         else {
4007                 if ((mask & GFP_NOFS) == GFP_NOFS)
4008                         mask = GFP_NOFS;
4009                 /*
4010                  * at this point we can safely clear everything except the
4011                  * locked bit and the nodatasum bit
4012                  */
4013                 ret = clear_extent_bit(tree, start, end,
4014                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4015                                  0, 0, NULL, mask);
4016
4017                 /* if clear_extent_bit failed for enomem reasons,
4018                  * we can't allow the release to continue.
4019                  */
4020                 if (ret < 0)
4021                         ret = 0;
4022                 else
4023                         ret = 1;
4024         }
4025         return ret;
4026 }
4027
4028 /*
4029  * a helper for releasepage.  As long as there are no locked extents
4030  * in the range corresponding to the page, both state records and extent
4031  * map records are removed
4032  */
4033 int try_release_extent_mapping(struct extent_map_tree *map,
4034                                struct extent_io_tree *tree, struct page *page,
4035                                gfp_t mask)
4036 {
4037         struct extent_map *em;
4038         u64 start = page_offset(page);
4039         u64 end = start + PAGE_CACHE_SIZE - 1;
4040
4041         if ((mask & __GFP_WAIT) &&
4042             page->mapping->host->i_size > 16 * 1024 * 1024) {
4043                 u64 len;
4044                 while (start <= end) {
4045                         len = end - start + 1;
4046                         write_lock(&map->lock);
4047                         em = lookup_extent_mapping(map, start, len);
4048                         if (!em) {
4049                                 write_unlock(&map->lock);
4050                                 break;
4051                         }
4052                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4053                             em->start != start) {
4054                                 write_unlock(&map->lock);
4055                                 free_extent_map(em);
4056                                 break;
4057                         }
4058                         if (!test_range_bit(tree, em->start,
4059                                             extent_map_end(em) - 1,
4060                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4061                                             0, NULL)) {
4062                                 remove_extent_mapping(map, em);
4063                                 /* once for the rb tree */
4064                                 free_extent_map(em);
4065                         }
4066                         start = extent_map_end(em);
4067                         write_unlock(&map->lock);
4068
4069                         /* once for us */
4070                         free_extent_map(em);
4071                 }
4072         }
4073         return try_release_extent_state(map, tree, page, mask);
4074 }
4075
4076 /*
4077  * helper function for fiemap, which doesn't want to see any holes.
4078  * This maps until we find something past 'last'
4079  */
4080 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4081                                                 u64 offset,
4082                                                 u64 last,
4083                                                 get_extent_t *get_extent)
4084 {
4085         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4086         struct extent_map *em;
4087         u64 len;
4088
4089         if (offset >= last)
4090                 return NULL;
4091
4092         while (1) {
4093                 len = last - offset;
4094                 if (len == 0)
4095                         break;
4096                 len = ALIGN(len, sectorsize);
4097                 em = get_extent(inode, NULL, 0, offset, len, 0);
4098                 if (IS_ERR_OR_NULL(em))
4099                         return em;
4100
4101                 /* if this isn't a hole return it */
4102                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4103                     em->block_start != EXTENT_MAP_HOLE) {
4104                         return em;
4105                 }
4106
4107                 /* this is a hole, advance to the next extent */
4108                 offset = extent_map_end(em);
4109                 free_extent_map(em);
4110                 if (offset >= last)
4111                         break;
4112         }
4113         return NULL;
4114 }
4115
4116 static noinline int count_ext_ref(u64 inum, u64 offset, u64 root_id, void *ctx)
4117 {
4118         unsigned long cnt = *((unsigned long *)ctx);
4119
4120         cnt++;
4121         *((unsigned long *)ctx) = cnt;
4122
4123         /* Now we're sure that the extent is shared. */
4124         if (cnt > 1)
4125                 return 1;
4126         return 0;
4127 }
4128
4129 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4130                 __u64 start, __u64 len, get_extent_t *get_extent)
4131 {
4132         int ret = 0;
4133         u64 off = start;
4134         u64 max = start + len;
4135         u32 flags = 0;
4136         u32 found_type;
4137         u64 last;
4138         u64 last_for_get_extent = 0;
4139         u64 disko = 0;
4140         u64 isize = i_size_read(inode);
4141         struct btrfs_key found_key;
4142         struct extent_map *em = NULL;
4143         struct extent_state *cached_state = NULL;
4144         struct btrfs_path *path;
4145         int end = 0;
4146         u64 em_start = 0;
4147         u64 em_len = 0;
4148         u64 em_end = 0;
4149
4150         if (len == 0)
4151                 return -EINVAL;
4152
4153         path = btrfs_alloc_path();
4154         if (!path)
4155                 return -ENOMEM;
4156         path->leave_spinning = 1;
4157
4158         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
4159         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
4160
4161         /*
4162          * lookup the last file extent.  We're not using i_size here
4163          * because there might be preallocation past i_size
4164          */
4165         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
4166                                        path, btrfs_ino(inode), -1, 0);
4167         if (ret < 0) {
4168                 btrfs_free_path(path);
4169                 return ret;
4170         }
4171         WARN_ON(!ret);
4172         path->slots[0]--;
4173         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4174         found_type = btrfs_key_type(&found_key);
4175
4176         /* No extents, but there might be delalloc bits */
4177         if (found_key.objectid != btrfs_ino(inode) ||
4178             found_type != BTRFS_EXTENT_DATA_KEY) {
4179                 /* have to trust i_size as the end */
4180                 last = (u64)-1;
4181                 last_for_get_extent = isize;
4182         } else {
4183                 /*
4184                  * remember the start of the last extent.  There are a
4185                  * bunch of different factors that go into the length of the
4186                  * extent, so its much less complex to remember where it started
4187                  */
4188                 last = found_key.offset;
4189                 last_for_get_extent = last + 1;
4190         }
4191         btrfs_release_path(path);
4192
4193         /*
4194          * we might have some extents allocated but more delalloc past those
4195          * extents.  so, we trust isize unless the start of the last extent is
4196          * beyond isize
4197          */
4198         if (last < isize) {
4199                 last = (u64)-1;
4200                 last_for_get_extent = isize;
4201         }
4202
4203         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4204                          &cached_state);
4205
4206         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4207                                    get_extent);
4208         if (!em)
4209                 goto out;
4210         if (IS_ERR(em)) {
4211                 ret = PTR_ERR(em);
4212                 goto out;
4213         }
4214
4215         while (!end) {
4216                 u64 offset_in_extent = 0;
4217
4218                 /* break if the extent we found is outside the range */
4219                 if (em->start >= max || extent_map_end(em) < off)
4220                         break;
4221
4222                 /*
4223                  * get_extent may return an extent that starts before our
4224                  * requested range.  We have to make sure the ranges
4225                  * we return to fiemap always move forward and don't
4226                  * overlap, so adjust the offsets here
4227                  */
4228                 em_start = max(em->start, off);
4229
4230                 /*
4231                  * record the offset from the start of the extent
4232                  * for adjusting the disk offset below.  Only do this if the
4233                  * extent isn't compressed since our in ram offset may be past
4234                  * what we have actually allocated on disk.
4235                  */
4236                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4237                         offset_in_extent = em_start - em->start;
4238                 em_end = extent_map_end(em);
4239                 em_len = em_end - em_start;
4240                 disko = 0;
4241                 flags = 0;
4242
4243                 /*
4244                  * bump off for our next call to get_extent
4245                  */
4246                 off = extent_map_end(em);
4247                 if (off >= max)
4248                         end = 1;
4249
4250                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4251                         end = 1;
4252                         flags |= FIEMAP_EXTENT_LAST;
4253                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4254                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4255                                   FIEMAP_EXTENT_NOT_ALIGNED);
4256                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4257                         flags |= (FIEMAP_EXTENT_DELALLOC |
4258                                   FIEMAP_EXTENT_UNKNOWN);
4259                 } else {
4260                         unsigned long ref_cnt = 0;
4261
4262                         disko = em->block_start + offset_in_extent;
4263
4264                         /*
4265                          * As btrfs supports shared space, this information
4266                          * can be exported to userspace tools via
4267                          * flag FIEMAP_EXTENT_SHARED.
4268                          */
4269                         ret = iterate_inodes_from_logical(
4270                                         em->block_start,
4271                                         BTRFS_I(inode)->root->fs_info,
4272                                         path, count_ext_ref, &ref_cnt);
4273                         if (ret < 0 && ret != -ENOENT)
4274                                 goto out_free;
4275
4276                         if (ref_cnt > 1)
4277                                 flags |= FIEMAP_EXTENT_SHARED;
4278                 }
4279                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4280                         flags |= FIEMAP_EXTENT_ENCODED;
4281
4282                 free_extent_map(em);
4283                 em = NULL;
4284                 if ((em_start >= last) || em_len == (u64)-1 ||
4285                    (last == (u64)-1 && isize <= em_end)) {
4286                         flags |= FIEMAP_EXTENT_LAST;
4287                         end = 1;
4288                 }
4289
4290                 /* now scan forward to see if this is really the last extent. */
4291                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4292                                            get_extent);
4293                 if (IS_ERR(em)) {
4294                         ret = PTR_ERR(em);
4295                         goto out;
4296                 }
4297                 if (!em) {
4298                         flags |= FIEMAP_EXTENT_LAST;
4299                         end = 1;
4300                 }
4301                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4302                                               em_len, flags);
4303                 if (ret)
4304                         goto out_free;
4305         }
4306 out_free:
4307         free_extent_map(em);
4308 out:
4309         btrfs_free_path(path);
4310         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4311                              &cached_state, GFP_NOFS);
4312         return ret;
4313 }
4314
4315 static void __free_extent_buffer(struct extent_buffer *eb)
4316 {
4317         btrfs_leak_debug_del(&eb->leak_list);
4318         kmem_cache_free(extent_buffer_cache, eb);
4319 }
4320
4321 int extent_buffer_under_io(struct extent_buffer *eb)
4322 {
4323         return (atomic_read(&eb->io_pages) ||
4324                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4325                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4326 }
4327
4328 /*
4329  * Helper for releasing extent buffer page.
4330  */
4331 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
4332                                                 unsigned long start_idx)
4333 {
4334         unsigned long index;
4335         unsigned long num_pages;
4336         struct page *page;
4337         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4338
4339         BUG_ON(extent_buffer_under_io(eb));
4340
4341         num_pages = num_extent_pages(eb->start, eb->len);
4342         index = start_idx + num_pages;
4343         if (start_idx >= index)
4344                 return;
4345
4346         do {
4347                 index--;
4348                 page = extent_buffer_page(eb, index);
4349                 if (page && mapped) {
4350                         spin_lock(&page->mapping->private_lock);
4351                         /*
4352                          * We do this since we'll remove the pages after we've
4353                          * removed the eb from the radix tree, so we could race
4354                          * and have this page now attached to the new eb.  So
4355                          * only clear page_private if it's still connected to
4356                          * this eb.
4357                          */
4358                         if (PagePrivate(page) &&
4359                             page->private == (unsigned long)eb) {
4360                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4361                                 BUG_ON(PageDirty(page));
4362                                 BUG_ON(PageWriteback(page));
4363                                 /*
4364                                  * We need to make sure we haven't be attached
4365                                  * to a new eb.
4366                                  */
4367                                 ClearPagePrivate(page);
4368                                 set_page_private(page, 0);
4369                                 /* One for the page private */
4370                                 page_cache_release(page);
4371                         }
4372                         spin_unlock(&page->mapping->private_lock);
4373
4374                 }
4375                 if (page) {
4376                         /* One for when we alloced the page */
4377                         page_cache_release(page);
4378                 }
4379         } while (index != start_idx);
4380 }
4381
4382 /*
4383  * Helper for releasing the extent buffer.
4384  */
4385 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4386 {
4387         btrfs_release_extent_buffer_page(eb, 0);
4388         __free_extent_buffer(eb);
4389 }
4390
4391 static struct extent_buffer *
4392 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4393                       unsigned long len, gfp_t mask)
4394 {
4395         struct extent_buffer *eb = NULL;
4396
4397         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
4398         if (eb == NULL)
4399                 return NULL;
4400         eb->start = start;
4401         eb->len = len;
4402         eb->fs_info = fs_info;
4403         eb->bflags = 0;
4404         rwlock_init(&eb->lock);
4405         atomic_set(&eb->write_locks, 0);
4406         atomic_set(&eb->read_locks, 0);
4407         atomic_set(&eb->blocking_readers, 0);
4408         atomic_set(&eb->blocking_writers, 0);
4409         atomic_set(&eb->spinning_readers, 0);
4410         atomic_set(&eb->spinning_writers, 0);
4411         eb->lock_nested = 0;
4412         init_waitqueue_head(&eb->write_lock_wq);
4413         init_waitqueue_head(&eb->read_lock_wq);
4414
4415         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4416
4417         spin_lock_init(&eb->refs_lock);
4418         atomic_set(&eb->refs, 1);
4419         atomic_set(&eb->io_pages, 0);
4420
4421         /*
4422          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4423          */
4424         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4425                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4426         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4427
4428         return eb;
4429 }
4430
4431 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4432 {
4433         unsigned long i;
4434         struct page *p;
4435         struct extent_buffer *new;
4436         unsigned long num_pages = num_extent_pages(src->start, src->len);
4437
4438         new = __alloc_extent_buffer(NULL, src->start, src->len, GFP_NOFS);
4439         if (new == NULL)
4440                 return NULL;
4441
4442         for (i = 0; i < num_pages; i++) {
4443                 p = alloc_page(GFP_NOFS);
4444                 if (!p) {
4445                         btrfs_release_extent_buffer(new);
4446                         return NULL;
4447                 }
4448                 attach_extent_buffer_page(new, p);
4449                 WARN_ON(PageDirty(p));
4450                 SetPageUptodate(p);
4451                 new->pages[i] = p;
4452         }
4453
4454         copy_extent_buffer(new, src, 0, 0, src->len);
4455         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4456         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4457
4458         return new;
4459 }
4460
4461 struct extent_buffer *alloc_dummy_extent_buffer(u64 start, unsigned long len)
4462 {
4463         struct extent_buffer *eb;
4464         unsigned long num_pages = num_extent_pages(0, len);
4465         unsigned long i;
4466
4467         eb = __alloc_extent_buffer(NULL, start, len, GFP_NOFS);
4468         if (!eb)
4469                 return NULL;
4470
4471         for (i = 0; i < num_pages; i++) {
4472                 eb->pages[i] = alloc_page(GFP_NOFS);
4473                 if (!eb->pages[i])
4474                         goto err;
4475         }
4476         set_extent_buffer_uptodate(eb);
4477         btrfs_set_header_nritems(eb, 0);
4478         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4479
4480         return eb;
4481 err:
4482         for (; i > 0; i--)
4483                 __free_page(eb->pages[i - 1]);
4484         __free_extent_buffer(eb);
4485         return NULL;
4486 }
4487
4488 static void check_buffer_tree_ref(struct extent_buffer *eb)
4489 {
4490         int refs;
4491         /* the ref bit is tricky.  We have to make sure it is set
4492          * if we have the buffer dirty.   Otherwise the
4493          * code to free a buffer can end up dropping a dirty
4494          * page
4495          *
4496          * Once the ref bit is set, it won't go away while the
4497          * buffer is dirty or in writeback, and it also won't
4498          * go away while we have the reference count on the
4499          * eb bumped.
4500          *
4501          * We can't just set the ref bit without bumping the
4502          * ref on the eb because free_extent_buffer might
4503          * see the ref bit and try to clear it.  If this happens
4504          * free_extent_buffer might end up dropping our original
4505          * ref by mistake and freeing the page before we are able
4506          * to add one more ref.
4507          *
4508          * So bump the ref count first, then set the bit.  If someone
4509          * beat us to it, drop the ref we added.
4510          */
4511         refs = atomic_read(&eb->refs);
4512         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4513                 return;
4514
4515         spin_lock(&eb->refs_lock);
4516         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4517                 atomic_inc(&eb->refs);
4518         spin_unlock(&eb->refs_lock);
4519 }
4520
4521 static void mark_extent_buffer_accessed(struct extent_buffer *eb)
4522 {
4523         unsigned long num_pages, i;
4524
4525         check_buffer_tree_ref(eb);
4526
4527         num_pages = num_extent_pages(eb->start, eb->len);
4528         for (i = 0; i < num_pages; i++) {
4529                 struct page *p = extent_buffer_page(eb, i);
4530                 mark_page_accessed(p);
4531         }
4532 }
4533
4534 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4535                                          u64 start)
4536 {
4537         struct extent_buffer *eb;
4538
4539         rcu_read_lock();
4540         eb = radix_tree_lookup(&fs_info->buffer_radix,
4541                                start >> PAGE_CACHE_SHIFT);
4542         if (eb && atomic_inc_not_zero(&eb->refs)) {
4543                 rcu_read_unlock();
4544                 mark_extent_buffer_accessed(eb);
4545                 return eb;
4546         }
4547         rcu_read_unlock();
4548
4549         return NULL;
4550 }
4551
4552 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4553 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4554                                                u64 start, unsigned long len)
4555 {
4556         struct extent_buffer *eb, *exists = NULL;
4557         int ret;
4558
4559         eb = find_extent_buffer(fs_info, start);
4560         if (eb)
4561                 return eb;
4562         eb = alloc_dummy_extent_buffer(start, len);
4563         if (!eb)
4564                 return NULL;
4565         eb->fs_info = fs_info;
4566 again:
4567         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4568         if (ret)
4569                 goto free_eb;
4570         spin_lock(&fs_info->buffer_lock);
4571         ret = radix_tree_insert(&fs_info->buffer_radix,
4572                                 start >> PAGE_CACHE_SHIFT, eb);
4573         spin_unlock(&fs_info->buffer_lock);
4574         radix_tree_preload_end();
4575         if (ret == -EEXIST) {
4576                 exists = find_extent_buffer(fs_info, start);
4577                 if (exists)
4578                         goto free_eb;
4579                 else
4580                         goto again;
4581         }
4582         check_buffer_tree_ref(eb);
4583         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4584
4585         /*
4586          * We will free dummy extent buffer's if they come into
4587          * free_extent_buffer with a ref count of 2, but if we are using this we
4588          * want the buffers to stay in memory until we're done with them, so
4589          * bump the ref count again.
4590          */
4591         atomic_inc(&eb->refs);
4592         return eb;
4593 free_eb:
4594         btrfs_release_extent_buffer(eb);
4595         return exists;
4596 }
4597 #endif
4598
4599 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4600                                           u64 start, unsigned long len)
4601 {
4602         unsigned long num_pages = num_extent_pages(start, len);
4603         unsigned long i;
4604         unsigned long index = start >> PAGE_CACHE_SHIFT;
4605         struct extent_buffer *eb;
4606         struct extent_buffer *exists = NULL;
4607         struct page *p;
4608         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4609         int uptodate = 1;
4610         int ret;
4611
4612         eb = find_extent_buffer(fs_info, start);
4613         if (eb)
4614                 return eb;
4615
4616         eb = __alloc_extent_buffer(fs_info, start, len, GFP_NOFS);
4617         if (!eb)
4618                 return NULL;
4619
4620         for (i = 0; i < num_pages; i++, index++) {
4621                 p = find_or_create_page(mapping, index, GFP_NOFS);
4622                 if (!p)
4623                         goto free_eb;
4624
4625                 spin_lock(&mapping->private_lock);
4626                 if (PagePrivate(p)) {
4627                         /*
4628                          * We could have already allocated an eb for this page
4629                          * and attached one so lets see if we can get a ref on
4630                          * the existing eb, and if we can we know it's good and
4631                          * we can just return that one, else we know we can just
4632                          * overwrite page->private.
4633                          */
4634                         exists = (struct extent_buffer *)p->private;
4635                         if (atomic_inc_not_zero(&exists->refs)) {
4636                                 spin_unlock(&mapping->private_lock);
4637                                 unlock_page(p);
4638                                 page_cache_release(p);
4639                                 mark_extent_buffer_accessed(exists);
4640                                 goto free_eb;
4641                         }
4642
4643                         /*
4644                          * Do this so attach doesn't complain and we need to
4645                          * drop the ref the old guy had.
4646                          */
4647                         ClearPagePrivate(p);
4648                         WARN_ON(PageDirty(p));
4649                         page_cache_release(p);
4650                 }
4651                 attach_extent_buffer_page(eb, p);
4652                 spin_unlock(&mapping->private_lock);
4653                 WARN_ON(PageDirty(p));
4654                 mark_page_accessed(p);
4655                 eb->pages[i] = p;
4656                 if (!PageUptodate(p))
4657                         uptodate = 0;
4658
4659                 /*
4660                  * see below about how we avoid a nasty race with release page
4661                  * and why we unlock later
4662                  */
4663         }
4664         if (uptodate)
4665                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4666 again:
4667         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4668         if (ret)
4669                 goto free_eb;
4670
4671         spin_lock(&fs_info->buffer_lock);
4672         ret = radix_tree_insert(&fs_info->buffer_radix,
4673                                 start >> PAGE_CACHE_SHIFT, eb);
4674         spin_unlock(&fs_info->buffer_lock);
4675         radix_tree_preload_end();
4676         if (ret == -EEXIST) {
4677                 exists = find_extent_buffer(fs_info, start);
4678                 if (exists)
4679                         goto free_eb;
4680                 else
4681                         goto again;
4682         }
4683         /* add one reference for the tree */
4684         check_buffer_tree_ref(eb);
4685         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4686
4687         /*
4688          * there is a race where release page may have
4689          * tried to find this extent buffer in the radix
4690          * but failed.  It will tell the VM it is safe to
4691          * reclaim the, and it will clear the page private bit.
4692          * We must make sure to set the page private bit properly
4693          * after the extent buffer is in the radix tree so
4694          * it doesn't get lost
4695          */
4696         SetPageChecked(eb->pages[0]);
4697         for (i = 1; i < num_pages; i++) {
4698                 p = extent_buffer_page(eb, i);
4699                 ClearPageChecked(p);
4700                 unlock_page(p);
4701         }
4702         unlock_page(eb->pages[0]);
4703         return eb;
4704
4705 free_eb:
4706         for (i = 0; i < num_pages; i++) {
4707                 if (eb->pages[i])
4708                         unlock_page(eb->pages[i]);
4709         }
4710
4711         WARN_ON(!atomic_dec_and_test(&eb->refs));
4712         btrfs_release_extent_buffer(eb);
4713         return exists;
4714 }
4715
4716 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4717 {
4718         struct extent_buffer *eb =
4719                         container_of(head, struct extent_buffer, rcu_head);
4720
4721         __free_extent_buffer(eb);
4722 }
4723
4724 /* Expects to have eb->eb_lock already held */
4725 static int release_extent_buffer(struct extent_buffer *eb)
4726 {
4727         WARN_ON(atomic_read(&eb->refs) == 0);
4728         if (atomic_dec_and_test(&eb->refs)) {
4729                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
4730                         struct btrfs_fs_info *fs_info = eb->fs_info;
4731
4732                         spin_unlock(&eb->refs_lock);
4733
4734                         spin_lock(&fs_info->buffer_lock);
4735                         radix_tree_delete(&fs_info->buffer_radix,
4736                                           eb->start >> PAGE_CACHE_SHIFT);
4737                         spin_unlock(&fs_info->buffer_lock);
4738                 } else {
4739                         spin_unlock(&eb->refs_lock);
4740                 }
4741
4742                 /* Should be safe to release our pages at this point */
4743                 btrfs_release_extent_buffer_page(eb, 0);
4744                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4745                 return 1;
4746         }
4747         spin_unlock(&eb->refs_lock);
4748
4749         return 0;
4750 }
4751
4752 void free_extent_buffer(struct extent_buffer *eb)
4753 {
4754         int refs;
4755         int old;
4756         if (!eb)
4757                 return;
4758
4759         while (1) {
4760                 refs = atomic_read(&eb->refs);
4761                 if (refs <= 3)
4762                         break;
4763                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
4764                 if (old == refs)
4765                         return;
4766         }
4767
4768         spin_lock(&eb->refs_lock);
4769         if (atomic_read(&eb->refs) == 2 &&
4770             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
4771                 atomic_dec(&eb->refs);
4772
4773         if (atomic_read(&eb->refs) == 2 &&
4774             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4775             !extent_buffer_under_io(eb) &&
4776             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4777                 atomic_dec(&eb->refs);
4778
4779         /*
4780          * I know this is terrible, but it's temporary until we stop tracking
4781          * the uptodate bits and such for the extent buffers.
4782          */
4783         release_extent_buffer(eb);
4784 }
4785
4786 void free_extent_buffer_stale(struct extent_buffer *eb)
4787 {
4788         if (!eb)
4789                 return;
4790
4791         spin_lock(&eb->refs_lock);
4792         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4793
4794         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4795             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4796                 atomic_dec(&eb->refs);
4797         release_extent_buffer(eb);
4798 }
4799
4800 void clear_extent_buffer_dirty(struct extent_buffer *eb)
4801 {
4802         unsigned long i;
4803         unsigned long num_pages;
4804         struct page *page;
4805
4806         num_pages = num_extent_pages(eb->start, eb->len);
4807
4808         for (i = 0; i < num_pages; i++) {
4809                 page = extent_buffer_page(eb, i);
4810                 if (!PageDirty(page))
4811                         continue;
4812
4813                 lock_page(page);
4814                 WARN_ON(!PagePrivate(page));
4815
4816                 clear_page_dirty_for_io(page);
4817                 spin_lock_irq(&page->mapping->tree_lock);
4818                 if (!PageDirty(page)) {
4819                         radix_tree_tag_clear(&page->mapping->page_tree,
4820                                                 page_index(page),
4821                                                 PAGECACHE_TAG_DIRTY);
4822                 }
4823                 spin_unlock_irq(&page->mapping->tree_lock);
4824                 ClearPageError(page);
4825                 unlock_page(page);
4826         }
4827         WARN_ON(atomic_read(&eb->refs) == 0);
4828 }
4829
4830 int set_extent_buffer_dirty(struct extent_buffer *eb)
4831 {
4832         unsigned long i;
4833         unsigned long num_pages;
4834         int was_dirty = 0;
4835
4836         check_buffer_tree_ref(eb);
4837
4838         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4839
4840         num_pages = num_extent_pages(eb->start, eb->len);
4841         WARN_ON(atomic_read(&eb->refs) == 0);
4842         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4843
4844         for (i = 0; i < num_pages; i++)
4845                 set_page_dirty(extent_buffer_page(eb, i));
4846         return was_dirty;
4847 }
4848
4849 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4850 {
4851         unsigned long i;
4852         struct page *page;
4853         unsigned long num_pages;
4854
4855         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4856         num_pages = num_extent_pages(eb->start, eb->len);
4857         for (i = 0; i < num_pages; i++) {
4858                 page = extent_buffer_page(eb, i);
4859                 if (page)
4860                         ClearPageUptodate(page);
4861         }
4862         return 0;
4863 }
4864
4865 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4866 {
4867         unsigned long i;
4868         struct page *page;
4869         unsigned long num_pages;
4870
4871         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4872         num_pages = num_extent_pages(eb->start, eb->len);
4873         for (i = 0; i < num_pages; i++) {
4874                 page = extent_buffer_page(eb, i);
4875                 SetPageUptodate(page);
4876         }
4877         return 0;
4878 }
4879
4880 int extent_buffer_uptodate(struct extent_buffer *eb)
4881 {
4882         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4883 }
4884
4885 int read_extent_buffer_pages(struct extent_io_tree *tree,
4886                              struct extent_buffer *eb, u64 start, int wait,
4887                              get_extent_t *get_extent, int mirror_num)
4888 {
4889         unsigned long i;
4890         unsigned long start_i;
4891         struct page *page;
4892         int err;
4893         int ret = 0;
4894         int locked_pages = 0;
4895         int all_uptodate = 1;
4896         unsigned long num_pages;
4897         unsigned long num_reads = 0;
4898         struct bio *bio = NULL;
4899         unsigned long bio_flags = 0;
4900
4901         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4902                 return 0;
4903
4904         if (start) {
4905                 WARN_ON(start < eb->start);
4906                 start_i = (start >> PAGE_CACHE_SHIFT) -
4907                         (eb->start >> PAGE_CACHE_SHIFT);
4908         } else {
4909                 start_i = 0;
4910         }
4911
4912         num_pages = num_extent_pages(eb->start, eb->len);
4913         for (i = start_i; i < num_pages; i++) {
4914                 page = extent_buffer_page(eb, i);
4915                 if (wait == WAIT_NONE) {
4916                         if (!trylock_page(page))
4917                                 goto unlock_exit;
4918                 } else {
4919                         lock_page(page);
4920                 }
4921                 locked_pages++;
4922                 if (!PageUptodate(page)) {
4923                         num_reads++;
4924                         all_uptodate = 0;
4925                 }
4926         }
4927         if (all_uptodate) {
4928                 if (start_i == 0)
4929                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4930                 goto unlock_exit;
4931         }
4932
4933         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4934         eb->read_mirror = 0;
4935         atomic_set(&eb->io_pages, num_reads);
4936         for (i = start_i; i < num_pages; i++) {
4937                 page = extent_buffer_page(eb, i);
4938                 if (!PageUptodate(page)) {
4939                         ClearPageError(page);
4940                         err = __extent_read_full_page(tree, page,
4941                                                       get_extent, &bio,
4942                                                       mirror_num, &bio_flags,
4943                                                       READ | REQ_META);
4944                         if (err)
4945                                 ret = err;
4946                 } else {
4947                         unlock_page(page);
4948                 }
4949         }
4950
4951         if (bio) {
4952                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
4953                                      bio_flags);
4954                 if (err)
4955                         return err;
4956         }
4957
4958         if (ret || wait != WAIT_COMPLETE)
4959                 return ret;
4960
4961         for (i = start_i; i < num_pages; i++) {
4962                 page = extent_buffer_page(eb, i);
4963                 wait_on_page_locked(page);
4964                 if (!PageUptodate(page))
4965                         ret = -EIO;
4966         }
4967
4968         return ret;
4969
4970 unlock_exit:
4971         i = start_i;
4972         while (locked_pages > 0) {
4973                 page = extent_buffer_page(eb, i);
4974                 i++;
4975                 unlock_page(page);
4976                 locked_pages--;
4977         }
4978         return ret;
4979 }
4980
4981 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4982                         unsigned long start,
4983                         unsigned long len)
4984 {
4985         size_t cur;
4986         size_t offset;
4987         struct page *page;
4988         char *kaddr;
4989         char *dst = (char *)dstv;
4990         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4991         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4992
4993         WARN_ON(start > eb->len);
4994         WARN_ON(start + len > eb->start + eb->len);
4995
4996         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
4997
4998         while (len > 0) {
4999                 page = extent_buffer_page(eb, i);
5000
5001                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5002                 kaddr = page_address(page);
5003                 memcpy(dst, kaddr + offset, cur);
5004
5005                 dst += cur;
5006                 len -= cur;
5007                 offset = 0;
5008                 i++;
5009         }
5010 }
5011
5012 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5013                                unsigned long min_len, char **map,
5014                                unsigned long *map_start,
5015                                unsigned long *map_len)
5016 {
5017         size_t offset = start & (PAGE_CACHE_SIZE - 1);
5018         char *kaddr;
5019         struct page *p;
5020         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5021         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5022         unsigned long end_i = (start_offset + start + min_len - 1) >>
5023                 PAGE_CACHE_SHIFT;
5024
5025         if (i != end_i)
5026                 return -EINVAL;
5027
5028         if (i == 0) {
5029                 offset = start_offset;
5030                 *map_start = 0;
5031         } else {
5032                 offset = 0;
5033                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
5034         }
5035
5036         if (start + min_len > eb->len) {
5037                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
5038                        "wanted %lu %lu\n",
5039                        eb->start, eb->len, start, min_len);
5040                 return -EINVAL;
5041         }
5042
5043         p = extent_buffer_page(eb, i);
5044         kaddr = page_address(p);
5045         *map = kaddr + offset;
5046         *map_len = PAGE_CACHE_SIZE - offset;
5047         return 0;
5048 }
5049
5050 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5051                           unsigned long start,
5052                           unsigned long len)
5053 {
5054         size_t cur;
5055         size_t offset;
5056         struct page *page;
5057         char *kaddr;
5058         char *ptr = (char *)ptrv;
5059         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5060         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5061         int ret = 0;
5062
5063         WARN_ON(start > eb->len);
5064         WARN_ON(start + len > eb->start + eb->len);
5065
5066         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5067
5068         while (len > 0) {
5069                 page = extent_buffer_page(eb, i);
5070
5071                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5072
5073                 kaddr = page_address(page);
5074                 ret = memcmp(ptr, kaddr + offset, cur);
5075                 if (ret)
5076                         break;
5077
5078                 ptr += cur;
5079                 len -= cur;
5080                 offset = 0;
5081                 i++;
5082         }
5083         return ret;
5084 }
5085
5086 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5087                          unsigned long start, unsigned long len)
5088 {
5089         size_t cur;
5090         size_t offset;
5091         struct page *page;
5092         char *kaddr;
5093         char *src = (char *)srcv;
5094         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5095         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5096
5097         WARN_ON(start > eb->len);
5098         WARN_ON(start + len > eb->start + eb->len);
5099
5100         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5101
5102         while (len > 0) {
5103                 page = extent_buffer_page(eb, i);
5104                 WARN_ON(!PageUptodate(page));
5105
5106                 cur = min(len, PAGE_CACHE_SIZE - offset);
5107                 kaddr = page_address(page);
5108                 memcpy(kaddr + offset, src, cur);
5109
5110                 src += cur;
5111                 len -= cur;
5112                 offset = 0;
5113                 i++;
5114         }
5115 }
5116
5117 void memset_extent_buffer(struct extent_buffer *eb, char c,
5118                           unsigned long start, unsigned long len)
5119 {
5120         size_t cur;
5121         size_t offset;
5122         struct page *page;
5123         char *kaddr;
5124         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5125         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5126
5127         WARN_ON(start > eb->len);
5128         WARN_ON(start + len > eb->start + eb->len);
5129
5130         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5131
5132         while (len > 0) {
5133                 page = extent_buffer_page(eb, i);
5134                 WARN_ON(!PageUptodate(page));
5135
5136                 cur = min(len, PAGE_CACHE_SIZE - offset);
5137                 kaddr = page_address(page);
5138                 memset(kaddr + offset, c, cur);
5139
5140                 len -= cur;
5141                 offset = 0;
5142                 i++;
5143         }
5144 }
5145
5146 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5147                         unsigned long dst_offset, unsigned long src_offset,
5148                         unsigned long len)
5149 {
5150         u64 dst_len = dst->len;
5151         size_t cur;
5152         size_t offset;
5153         struct page *page;
5154         char *kaddr;
5155         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5156         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5157
5158         WARN_ON(src->len != dst_len);
5159
5160         offset = (start_offset + dst_offset) &
5161                 (PAGE_CACHE_SIZE - 1);
5162
5163         while (len > 0) {
5164                 page = extent_buffer_page(dst, i);
5165                 WARN_ON(!PageUptodate(page));
5166
5167                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5168
5169                 kaddr = page_address(page);
5170                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5171
5172                 src_offset += cur;
5173                 len -= cur;
5174                 offset = 0;
5175                 i++;
5176         }
5177 }
5178
5179 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5180 {
5181         unsigned long distance = (src > dst) ? src - dst : dst - src;
5182         return distance < len;
5183 }
5184
5185 static void copy_pages(struct page *dst_page, struct page *src_page,
5186                        unsigned long dst_off, unsigned long src_off,
5187                        unsigned long len)
5188 {
5189         char *dst_kaddr = page_address(dst_page);
5190         char *src_kaddr;
5191         int must_memmove = 0;
5192
5193         if (dst_page != src_page) {
5194                 src_kaddr = page_address(src_page);
5195         } else {
5196                 src_kaddr = dst_kaddr;
5197                 if (areas_overlap(src_off, dst_off, len))
5198                         must_memmove = 1;
5199         }
5200
5201         if (must_memmove)
5202                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5203         else
5204                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5205 }
5206
5207 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5208                            unsigned long src_offset, unsigned long len)
5209 {
5210         size_t cur;
5211         size_t dst_off_in_page;
5212         size_t src_off_in_page;
5213         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5214         unsigned long dst_i;
5215         unsigned long src_i;
5216
5217         if (src_offset + len > dst->len) {
5218                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5219                        "len %lu dst len %lu\n", src_offset, len, dst->len);
5220                 BUG_ON(1);
5221         }
5222         if (dst_offset + len > dst->len) {
5223                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5224                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
5225                 BUG_ON(1);
5226         }
5227
5228         while (len > 0) {
5229                 dst_off_in_page = (start_offset + dst_offset) &
5230                         (PAGE_CACHE_SIZE - 1);
5231                 src_off_in_page = (start_offset + src_offset) &
5232                         (PAGE_CACHE_SIZE - 1);
5233
5234                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5235                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5236
5237                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5238                                                src_off_in_page));
5239                 cur = min_t(unsigned long, cur,
5240                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5241
5242                 copy_pages(extent_buffer_page(dst, dst_i),
5243                            extent_buffer_page(dst, src_i),
5244                            dst_off_in_page, src_off_in_page, cur);
5245
5246                 src_offset += cur;
5247                 dst_offset += cur;
5248                 len -= cur;
5249         }
5250 }
5251
5252 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5253                            unsigned long src_offset, unsigned long len)
5254 {
5255         size_t cur;
5256         size_t dst_off_in_page;
5257         size_t src_off_in_page;
5258         unsigned long dst_end = dst_offset + len - 1;
5259         unsigned long src_end = src_offset + len - 1;
5260         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5261         unsigned long dst_i;
5262         unsigned long src_i;
5263
5264         if (src_offset + len > dst->len) {
5265                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5266                        "len %lu len %lu\n", src_offset, len, dst->len);
5267                 BUG_ON(1);
5268         }
5269         if (dst_offset + len > dst->len) {
5270                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5271                        "len %lu len %lu\n", dst_offset, len, dst->len);
5272                 BUG_ON(1);
5273         }
5274         if (dst_offset < src_offset) {
5275                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5276                 return;
5277         }
5278         while (len > 0) {
5279                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5280                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5281
5282                 dst_off_in_page = (start_offset + dst_end) &
5283                         (PAGE_CACHE_SIZE - 1);
5284                 src_off_in_page = (start_offset + src_end) &
5285                         (PAGE_CACHE_SIZE - 1);
5286
5287                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5288                 cur = min(cur, dst_off_in_page + 1);
5289                 copy_pages(extent_buffer_page(dst, dst_i),
5290                            extent_buffer_page(dst, src_i),
5291                            dst_off_in_page - cur + 1,
5292                            src_off_in_page - cur + 1, cur);
5293
5294                 dst_end -= cur;
5295                 src_end -= cur;
5296                 len -= cur;
5297         }
5298 }
5299
5300 int try_release_extent_buffer(struct page *page)
5301 {
5302         struct extent_buffer *eb;
5303
5304         /*
5305          * We need to make sure noboody is attaching this page to an eb right
5306          * now.
5307          */
5308         spin_lock(&page->mapping->private_lock);
5309         if (!PagePrivate(page)) {
5310                 spin_unlock(&page->mapping->private_lock);
5311                 return 1;
5312         }
5313
5314         eb = (struct extent_buffer *)page->private;
5315         BUG_ON(!eb);
5316
5317         /*
5318          * This is a little awful but should be ok, we need to make sure that
5319          * the eb doesn't disappear out from under us while we're looking at
5320          * this page.
5321          */
5322         spin_lock(&eb->refs_lock);
5323         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5324                 spin_unlock(&eb->refs_lock);
5325                 spin_unlock(&page->mapping->private_lock);
5326                 return 0;
5327         }
5328         spin_unlock(&page->mapping->private_lock);
5329
5330         /*
5331          * If tree ref isn't set then we know the ref on this eb is a real ref,
5332          * so just return, this page will likely be freed soon anyway.
5333          */
5334         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5335                 spin_unlock(&eb->refs_lock);
5336                 return 0;
5337         }
5338
5339         return release_extent_buffer(eb);
5340 }