fs/btrfs/extent_io.c

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