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 static inline bool extent_state_in_tree(const struct extent_state *state)
  29 {
  30         return !RB_EMPTY_NODE(&state->rb_node);
  31 }
  32
  33 #ifdef CONFIG_BTRFS_DEBUG
  34 static LIST_HEAD(buffers);
  35 static LIST_HEAD(states);
  36
  37 static DEFINE_SPINLOCK(leak_lock);
  38
  39 static inline
  40 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  41 {
  42         unsigned long flags;
  43
  44         spin_lock_irqsave(&leak_lock, flags);
  45         list_add(new, head);
  46         spin_unlock_irqrestore(&leak_lock, flags);
  47 }
  48
  49 static inline
  50 void btrfs_leak_debug_del(struct list_head *entry)
  51 {
  52         unsigned long flags;
  53
  54         spin_lock_irqsave(&leak_lock, flags);
  55         list_del(entry);
  56         spin_unlock_irqrestore(&leak_lock, flags);
  57 }
  58
  59 static inline
  60 void btrfs_leak_debug_check(void)
  61 {
  62         struct extent_state *state;
  63         struct extent_buffer *eb;
  64
  65         while (!list_empty(&states)) {
  66                 state = list_entry(states.next, struct extent_state, leak_list);
  67                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  68                        state->start, state->end, state->state,
  69                        extent_state_in_tree(state),
  70                        atomic_read(&state->refs));
  71                 list_del(&state->leak_list);
  72                 kmem_cache_free(extent_state_cache, state);
  73         }
  74
  75         while (!list_empty(&buffers)) {
  76                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  77                 printk(KERN_ERR "BTRFS: buffer leak start %llu len %lu "
  78                        "refs %d\n",
  79                        eb->start, eb->len, atomic_read(&eb->refs));
  80                 list_del(&eb->leak_list);
  81                 kmem_cache_free(extent_buffer_cache, eb);
  82         }
  83 }
  84
  85 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
  86         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
  87 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
  88                 struct extent_io_tree *tree, u64 start, u64 end)
  89 {
  90         struct inode *inode;
  91         u64 isize;
  92
  93         if (!tree->mapping)
  94                 return;
  95
  96         inode = tree->mapping->host;
  97         isize = i_size_read(inode);
  98         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
  99                 printk_ratelimited(KERN_DEBUG
 100                     "BTRFS: %s: ino %llu isize %llu odd range [%llu,%llu]\n",
 101                                 caller, btrfs_ino(inode), isize, start, end);
 102         }
 103 }
 104 #else
 105 #define btrfs_leak_debug_add(new, head) do {} while (0)
 106 #define btrfs_leak_debug_del(entry)     do {} while (0)
 107 #define btrfs_leak_debug_check()        do {} while (0)
 108 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
 109 #endif
 110
 111 #define BUFFER_LRU_MAX 64
 112
 113 struct tree_entry {
 114         u64 start;
 115         u64 end;
 116         struct rb_node rb_node;
 117 };
 118
 119 struct extent_page_data {
 120         struct bio *bio;
 121         struct extent_io_tree *tree;
 122         get_extent_t *get_extent;
 123         unsigned long bio_flags;
 124
 125         /* tells writepage not to lock the state bits for this range
 126          * it still does the unlocking
 127          */
 128         unsigned int extent_locked:1;
 129
 130         /* tells the submit_bio code to use a WRITE_SYNC */
 131         unsigned int sync_io:1;
 132 };
 133
 134 static noinline void flush_write_bio(void *data);
 135 static inline struct btrfs_fs_info *
 136 tree_fs_info(struct extent_io_tree *tree)
 137 {
 138         if (!tree->mapping)
 139                 return NULL;
 140         return btrfs_sb(tree->mapping->host->i_sb);
 141 }
 142
 143 int __init extent_io_init(void)
 144 {
 145         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 146                         sizeof(struct extent_state), 0,
 147                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 148         if (!extent_state_cache)
 149                 return -ENOMEM;
 150
 151         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 152                         sizeof(struct extent_buffer), 0,
 153                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 154         if (!extent_buffer_cache)
 155                 goto free_state_cache;
 156
 157         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 158                                      offsetof(struct btrfs_io_bio, bio));
 159         if (!btrfs_bioset)
 160                 goto free_buffer_cache;
 161
 162         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 163                 goto free_bioset;
 164
 165         return 0;
 166
 167 free_bioset:
 168         bioset_free(btrfs_bioset);
 169         btrfs_bioset = NULL;
 170
 171 free_buffer_cache:
 172         kmem_cache_destroy(extent_buffer_cache);
 173         extent_buffer_cache = NULL;
 174
 175 free_state_cache:
 176         kmem_cache_destroy(extent_state_cache);
 177         extent_state_cache = NULL;
 178         return -ENOMEM;
 179 }
 180
 181 void extent_io_exit(void)
 182 {
 183         btrfs_leak_debug_check();
 184
 185         /*
 186          * Make sure all delayed rcu free are flushed before we
 187          * destroy caches.
 188          */
 189         rcu_barrier();
 190         if (extent_state_cache)
 191                 kmem_cache_destroy(extent_state_cache);
 192         if (extent_buffer_cache)
 193                 kmem_cache_destroy(extent_buffer_cache);
 194         if (btrfs_bioset)
 195                 bioset_free(btrfs_bioset);
 196 }
 197
 198 void extent_io_tree_init(struct extent_io_tree *tree,
 199                          struct address_space *mapping)
 200 {
 201         tree->state = RB_ROOT;
 202         tree->ops = NULL;
 203         tree->dirty_bytes = 0;
 204         spin_lock_init(&tree->lock);
 205         tree->mapping = mapping;
 206 }
 207
 208 static struct extent_state *alloc_extent_state(gfp_t mask)
 209 {
 210         struct extent_state *state;
 211
 212         state = kmem_cache_alloc(extent_state_cache, mask);
 213         if (!state)
 214                 return state;
 215         state->state = 0;
 216         state->private = 0;
 217         RB_CLEAR_NODE(&state->rb_node);
 218         btrfs_leak_debug_add(&state->leak_list, &states);
 219         atomic_set(&state->refs, 1);
 220         init_waitqueue_head(&state->wq);
 221         trace_alloc_extent_state(state, mask, _RET_IP_);
 222         return state;
 223 }
 224
 225 void free_extent_state(struct extent_state *state)
 226 {
 227         if (!state)
 228                 return;
 229         if (atomic_dec_and_test(&state->refs)) {
 230                 WARN_ON(extent_state_in_tree(state));
 231                 btrfs_leak_debug_del(&state->leak_list);
 232                 trace_free_extent_state(state, _RET_IP_);
 233                 kmem_cache_free(extent_state_cache, state);
 234         }
 235 }
 236
 237 static struct rb_node *tree_insert(struct rb_root *root,
 238                                    struct rb_node *search_start,
 239                                    u64 offset,
 240                                    struct rb_node *node,
 241                                    struct rb_node ***p_in,
 242                                    struct rb_node **parent_in)
 243 {
 244         struct rb_node **p;
 245         struct rb_node *parent = NULL;
 246         struct tree_entry *entry;
 247
 248         if (p_in && parent_in) {
 249                 p = *p_in;
 250                 parent = *parent_in;
 251                 goto do_insert;
 252         }
 253
 254         p = search_start ? &search_start : &root->rb_node;
 255         while (*p) {
 256                 parent = *p;
 257                 entry = rb_entry(parent, struct tree_entry, rb_node);
 258
 259                 if (offset < entry->start)
 260                         p = &(*p)->rb_left;
 261                 else if (offset > entry->end)
 262                         p = &(*p)->rb_right;
 263                 else
 264                         return parent;
 265         }
 266
 267 do_insert:
 268         rb_link_node(node, parent, p);
 269         rb_insert_color(node, root);
 270         return NULL;
 271 }
 272
 273 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 274                                       struct rb_node **prev_ret,
 275                                       struct rb_node **next_ret,
 276                                       struct rb_node ***p_ret,
 277                                       struct rb_node **parent_ret)
 278 {
 279         struct rb_root *root = &tree->state;
 280         struct rb_node **n = &root->rb_node;
 281         struct rb_node *prev = NULL;
 282         struct rb_node *orig_prev = NULL;
 283         struct tree_entry *entry;
 284         struct tree_entry *prev_entry = NULL;
 285
 286         while (*n) {
 287                 prev = *n;
 288                 entry = rb_entry(prev, struct tree_entry, rb_node);
 289                 prev_entry = entry;
 290
 291                 if (offset < entry->start)
 292                         n = &(*n)->rb_left;
 293                 else if (offset > entry->end)
 294                         n = &(*n)->rb_right;
 295                 else
 296                         return *n;
 297         }
 298
 299         if (p_ret)
 300                 *p_ret = n;
 301         if (parent_ret)
 302                 *parent_ret = prev;
 303
 304         if (prev_ret) {
 305                 orig_prev = prev;
 306                 while (prev && offset > prev_entry->end) {
 307                         prev = rb_next(prev);
 308                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 309                 }
 310                 *prev_ret = prev;
 311                 prev = orig_prev;
 312         }
 313
 314         if (next_ret) {
 315                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 316                 while (prev && offset < prev_entry->start) {
 317                         prev = rb_prev(prev);
 318                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 319                 }
 320                 *next_ret = prev;
 321         }
 322         return NULL;
 323 }
 324
 325 static inline struct rb_node *
 326 tree_search_for_insert(struct extent_io_tree *tree,
 327                        u64 offset,
 328                        struct rb_node ***p_ret,
 329                        struct rb_node **parent_ret)
 330 {
 331         struct rb_node *prev = NULL;
 332         struct rb_node *ret;
 333
 334         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 335         if (!ret)
 336                 return prev;
 337         return ret;
 338 }
 339
 340 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 341                                           u64 offset)
 342 {
 343         return tree_search_for_insert(tree, offset, NULL, NULL);
 344 }
 345
 346 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 347                      struct extent_state *other)
 348 {
 349         if (tree->ops && tree->ops->merge_extent_hook)
 350                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 351                                              other);
 352 }
 353
 354 /*
 355  * utility function to look for merge candidates inside a given range.
 356  * Any extents with matching state are merged together into a single
 357  * extent in the tree.  Extents with EXTENT_IO in their state field
 358  * are not merged because the end_io handlers need to be able to do
 359  * operations on them without sleeping (or doing allocations/splits).
 360  *
 361  * This should be called with the tree lock held.
 362  */
 363 static void merge_state(struct extent_io_tree *tree,
 364                         struct extent_state *state)
 365 {
 366         struct extent_state *other;
 367         struct rb_node *other_node;
 368
 369         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 370                 return;
 371
 372         other_node = rb_prev(&state->rb_node);
 373         if (other_node) {
 374                 other = rb_entry(other_node, struct extent_state, rb_node);
 375                 if (other->end == state->start - 1 &&
 376                     other->state == state->state) {
 377                         merge_cb(tree, state, other);
 378                         state->start = other->start;
 379                         rb_erase(&other->rb_node, &tree->state);
 380                         RB_CLEAR_NODE(&other->rb_node);
 381                         free_extent_state(other);
 382                 }
 383         }
 384         other_node = rb_next(&state->rb_node);
 385         if (other_node) {
 386                 other = rb_entry(other_node, struct extent_state, rb_node);
 387                 if (other->start == state->end + 1 &&
 388                     other->state == state->state) {
 389                         merge_cb(tree, state, other);
 390                         state->end = other->end;
 391                         rb_erase(&other->rb_node, &tree->state);
 392                         RB_CLEAR_NODE(&other->rb_node);
 393                         free_extent_state(other);
 394                 }
 395         }
 396 }
 397
 398 static void set_state_cb(struct extent_io_tree *tree,
 399                          struct extent_state *state, unsigned *bits)
 400 {
 401         if (tree->ops && tree->ops->set_bit_hook)
 402                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 403 }
 404
 405 static void clear_state_cb(struct extent_io_tree *tree,
 406                            struct extent_state *state, unsigned *bits)
 407 {
 408         if (tree->ops && tree->ops->clear_bit_hook)
 409                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
 410 }
 411
 412 static void set_state_bits(struct extent_io_tree *tree,
 413                            struct extent_state *state, unsigned *bits);
 414
 415 /*
 416  * insert an extent_state struct into the tree.  'bits' are set on the
 417  * struct before it is inserted.
 418  *
 419  * This may return -EEXIST if the extent is already there, in which case the
 420  * state struct is freed.
 421  *
 422  * The tree lock is not taken internally.  This is a utility function and
 423  * probably isn't what you want to call (see set/clear_extent_bit).
 424  */
 425 static int insert_state(struct extent_io_tree *tree,
 426                         struct extent_state *state, u64 start, u64 end,
 427                         struct rb_node ***p,
 428                         struct rb_node **parent,
 429                         unsigned *bits)
 430 {
 431         struct rb_node *node;
 432
 433         if (end < start)
 434                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 435                        end, start);
 436         state->start = start;
 437         state->end = end;
 438
 439         set_state_bits(tree, state, bits);
 440
 441         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 442         if (node) {
 443                 struct extent_state *found;
 444                 found = rb_entry(node, struct extent_state, rb_node);
 445                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
 446                        "%llu %llu\n",
 447                        found->start, found->end, start, end);
 448                 return -EEXIST;
 449         }
 450         merge_state(tree, state);
 451         return 0;
 452 }
 453
 454 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 455                      u64 split)
 456 {
 457         if (tree->ops && tree->ops->split_extent_hook)
 458                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 459 }
 460
 461 /*
 462  * split a given extent state struct in two, inserting the preallocated
 463  * struct 'prealloc' as the newly created second half.  'split' indicates an
 464  * offset inside 'orig' where it should be split.
 465  *
 466  * Before calling,
 467  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 468  * are two extent state structs in the tree:
 469  * prealloc: [orig->start, split - 1]
 470  * orig: [ split, orig->end ]
 471  *
 472  * The tree locks are not taken by this function. They need to be held
 473  * by the caller.
 474  */
 475 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 476                        struct extent_state *prealloc, u64 split)
 477 {
 478         struct rb_node *node;
 479
 480         split_cb(tree, orig, split);
 481
 482         prealloc->start = orig->start;
 483         prealloc->end = split - 1;
 484         prealloc->state = orig->state;
 485         orig->start = split;
 486
 487         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 488                            &prealloc->rb_node, NULL, NULL);
 489         if (node) {
 490                 free_extent_state(prealloc);
 491                 return -EEXIST;
 492         }
 493         return 0;
 494 }
 495
 496 static struct extent_state *next_state(struct extent_state *state)
 497 {
 498         struct rb_node *next = rb_next(&state->rb_node);
 499         if (next)
 500                 return rb_entry(next, struct extent_state, rb_node);
 501         else
 502                 return NULL;
 503 }
 504
 505 /*
 506  * utility function to clear some bits in an extent state struct.
 507  * it will optionally wake up any one waiting on this state (wake == 1).
 508  *
 509  * If no bits are set on the state struct after clearing things, the
 510  * struct is freed and removed from the tree
 511  */
 512 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 513                                             struct extent_state *state,
 514                                             unsigned *bits, int wake)
 515 {
 516         struct extent_state *next;
 517         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 518
 519         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 520                 u64 range = state->end - state->start + 1;
 521                 WARN_ON(range > tree->dirty_bytes);
 522                 tree->dirty_bytes -= range;
 523         }
 524         clear_state_cb(tree, state, bits);
 525         state->state &= ~bits_to_clear;
 526         if (wake)
 527                 wake_up(&state->wq);
 528         if (state->state == 0) {
 529                 next = next_state(state);
 530                 if (extent_state_in_tree(state)) {
 531                         rb_erase(&state->rb_node, &tree->state);
 532                         RB_CLEAR_NODE(&state->rb_node);
 533                         free_extent_state(state);
 534                 } else {
 535                         WARN_ON(1);
 536                 }
 537         } else {
 538                 merge_state(tree, state);
 539                 next = next_state(state);
 540         }
 541         return next;
 542 }
 543
 544 static struct extent_state *
 545 alloc_extent_state_atomic(struct extent_state *prealloc)
 546 {
 547         if (!prealloc)
 548                 prealloc = alloc_extent_state(GFP_ATOMIC);
 549
 550         return prealloc;
 551 }
 552
 553 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 554 {
 555         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
 556                     "Extent tree was modified by another "
 557                     "thread while locked.");
 558 }
 559
 560 /*
 561  * clear some bits on a range in the tree.  This may require splitting
 562  * or inserting elements in the tree, so the gfp mask is used to
 563  * indicate which allocations or sleeping are allowed.
 564  *
 565  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 566  * the given range from the tree regardless of state (ie for truncate).
 567  *
 568  * the range [start, end] is inclusive.
 569  *
 570  * This takes the tree lock, and returns 0 on success and < 0 on error.
 571  */
 572 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 573                      unsigned bits, int wake, int delete,
 574                      struct extent_state **cached_state,
 575                      gfp_t mask)
 576 {
 577         struct extent_state *state;
 578         struct extent_state *cached;
 579         struct extent_state *prealloc = NULL;
 580         struct rb_node *node;
 581         u64 last_end;
 582         int err;
 583         int clear = 0;
 584
 585         btrfs_debug_check_extent_io_range(tree, start, end);
 586
 587         if (bits & EXTENT_DELALLOC)
 588                 bits |= EXTENT_NORESERVE;
 589
 590         if (delete)
 591                 bits |= ~EXTENT_CTLBITS;
 592         bits |= EXTENT_FIRST_DELALLOC;
 593
 594         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 595                 clear = 1;
 596 again:
 597         if (!prealloc && (mask & __GFP_WAIT)) {
 598                 /*
 599                  * Don't care for allocation failure here because we might end
 600                  * up not needing the pre-allocated extent state at all, which
 601                  * is the case if we only have in the tree extent states that
 602                  * cover our input range and don't cover too any other range.
 603                  * If we end up needing a new extent state we allocate it later.
 604                  */
 605                 prealloc = alloc_extent_state(mask);
 606         }
 607
 608         spin_lock(&tree->lock);
 609         if (cached_state) {
 610                 cached = *cached_state;
 611
 612                 if (clear) {
 613                         *cached_state = NULL;
 614                         cached_state = NULL;
 615                 }
 616
 617                 if (cached && extent_state_in_tree(cached) &&
 618                     cached->start <= start && cached->end > start) {
 619                         if (clear)
 620                                 atomic_dec(&cached->refs);
 621                         state = cached;
 622                         goto hit_next;
 623                 }
 624                 if (clear)
 625                         free_extent_state(cached);
 626         }
 627         /*
 628          * this search will find the extents that end after
 629          * our range starts
 630          */
 631         node = tree_search(tree, start);
 632         if (!node)
 633                 goto out;
 634         state = rb_entry(node, struct extent_state, rb_node);
 635 hit_next:
 636         if (state->start > end)
 637                 goto out;
 638         WARN_ON(state->end < start);
 639         last_end = state->end;
 640
 641         /* the state doesn't have the wanted bits, go ahead */
 642         if (!(state->state & bits)) {
 643                 state = next_state(state);
 644                 goto next;
 645         }
 646
 647         /*
 648          *     | ---- desired range ---- |
 649          *  | state | or
 650          *  | ------------- state -------------- |
 651          *
 652          * We need to split the extent we found, and may flip
 653          * bits on second half.
 654          *
 655          * If the extent we found extends past our range, we
 656          * just split and search again.  It'll get split again
 657          * the next time though.
 658          *
 659          * If the extent we found is inside our range, we clear
 660          * the desired bit on it.
 661          */
 662
 663         if (state->start < start) {
 664                 prealloc = alloc_extent_state_atomic(prealloc);
 665                 BUG_ON(!prealloc);
 666                 err = split_state(tree, state, prealloc, start);
 667                 if (err)
 668                         extent_io_tree_panic(tree, err);
 669
 670                 prealloc = NULL;
 671                 if (err)
 672                         goto out;
 673                 if (state->end <= end) {
 674                         state = clear_state_bit(tree, state, &bits, wake);
 675                         goto next;
 676                 }
 677                 goto search_again;
 678         }
 679         /*
 680          * | ---- desired range ---- |
 681          *                        | state |
 682          * We need to split the extent, and clear the bit
 683          * on the first half
 684          */
 685         if (state->start <= end && state->end > end) {
 686                 prealloc = alloc_extent_state_atomic(prealloc);
 687                 BUG_ON(!prealloc);
 688                 err = split_state(tree, state, prealloc, end + 1);
 689                 if (err)
 690                         extent_io_tree_panic(tree, err);
 691
 692                 if (wake)
 693                         wake_up(&state->wq);
 694
 695                 clear_state_bit(tree, prealloc, &bits, wake);
 696
 697                 prealloc = NULL;
 698                 goto out;
 699         }
 700
 701         state = clear_state_bit(tree, state, &bits, wake);
 702 next:
 703         if (last_end == (u64)-1)
 704                 goto out;
 705         start = last_end + 1;
 706         if (start <= end && state && !need_resched())
 707                 goto hit_next;
 708         goto search_again;
 709
 710 out:
 711         spin_unlock(&tree->lock);
 712         if (prealloc)
 713                 free_extent_state(prealloc);
 714
 715         return 0;
 716
 717 search_again:
 718         if (start > end)
 719                 goto out;
 720         spin_unlock(&tree->lock);
 721         if (mask & __GFP_WAIT)
 722                 cond_resched();
 723         goto again;
 724 }
 725
 726 static void wait_on_state(struct extent_io_tree *tree,
 727                           struct extent_state *state)
 728                 __releases(tree->lock)
 729                 __acquires(tree->lock)
 730 {
 731         DEFINE_WAIT(wait);
 732         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 733         spin_unlock(&tree->lock);
 734         schedule();
 735         spin_lock(&tree->lock);
 736         finish_wait(&state->wq, &wait);
 737 }
 738
 739 /*
 740  * waits for one or more bits to clear on a range in the state tree.
 741  * The range [start, end] is inclusive.
 742  * The tree lock is taken by this function
 743  */
 744 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 745                             unsigned long bits)
 746 {
 747         struct extent_state *state;
 748         struct rb_node *node;
 749
 750         btrfs_debug_check_extent_io_range(tree, start, end);
 751
 752         spin_lock(&tree->lock);
 753 again:
 754         while (1) {
 755                 /*
 756                  * this search will find all the extents that end after
 757                  * our range starts
 758                  */
 759                 node = tree_search(tree, start);
 760 process_node:
 761                 if (!node)
 762                         break;
 763
 764                 state = rb_entry(node, struct extent_state, rb_node);
 765
 766                 if (state->start > end)
 767                         goto out;
 768
 769                 if (state->state & bits) {
 770                         start = state->start;
 771                         atomic_inc(&state->refs);
 772                         wait_on_state(tree, state);
 773                         free_extent_state(state);
 774                         goto again;
 775                 }
 776                 start = state->end + 1;
 777
 778                 if (start > end)
 779                         break;
 780
 781                 if (!cond_resched_lock(&tree->lock)) {
 782                         node = rb_next(node);
 783                         goto process_node;
 784                 }
 785         }
 786 out:
 787         spin_unlock(&tree->lock);
 788 }
 789
 790 static void set_state_bits(struct extent_io_tree *tree,
 791                            struct extent_state *state,
 792                            unsigned *bits)
 793 {
 794         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 795
 796         set_state_cb(tree, state, bits);
 797         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 798                 u64 range = state->end - state->start + 1;
 799                 tree->dirty_bytes += range;
 800         }
 801         state->state |= bits_to_set;
 802 }
 803
 804 static void cache_state_if_flags(struct extent_state *state,
 805                                  struct extent_state **cached_ptr,
 806                                  unsigned flags)
 807 {
 808         if (cached_ptr && !(*cached_ptr)) {
 809                 if (!flags || (state->state & flags)) {
 810                         *cached_ptr = state;
 811                         atomic_inc(&state->refs);
 812                 }
 813         }
 814 }
 815
 816 static void cache_state(struct extent_state *state,
 817                         struct extent_state **cached_ptr)
 818 {
 819         return cache_state_if_flags(state, cached_ptr,
 820                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 821 }
 822
 823 /*
 824  * set some bits on a range in the tree.  This may require allocations or
 825  * sleeping, so the gfp mask is used to indicate what is allowed.
 826  *
 827  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 828  * part of the range already has the desired bits set.  The start of the
 829  * existing range is returned in failed_start in this case.
 830  *
 831  * [start, end] is inclusive This takes the tree lock.
 832  */
 833
 834 static int __must_check
 835 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 836                  unsigned bits, unsigned exclusive_bits,
 837                  u64 *failed_start, struct extent_state **cached_state,
 838                  gfp_t mask)
 839 {
 840         struct extent_state *state;
 841         struct extent_state *prealloc = NULL;
 842         struct rb_node *node;
 843         struct rb_node **p;
 844         struct rb_node *parent;
 845         int err = 0;
 846         u64 last_start;
 847         u64 last_end;
 848
 849         btrfs_debug_check_extent_io_range(tree, start, end);
 850
 851         bits |= EXTENT_FIRST_DELALLOC;
 852 again:
 853         if (!prealloc && (mask & __GFP_WAIT)) {
 854                 prealloc = alloc_extent_state(mask);
 855                 BUG_ON(!prealloc);
 856         }
 857
 858         spin_lock(&tree->lock);
 859         if (cached_state && *cached_state) {
 860                 state = *cached_state;
 861                 if (state->start <= start && state->end > start &&
 862                     extent_state_in_tree(state)) {
 863                         node = &state->rb_node;
 864                         goto hit_next;
 865                 }
 866         }
 867         /*
 868          * this search will find all the extents that end after
 869          * our range starts.
 870          */
 871         node = tree_search_for_insert(tree, start, &p, &parent);
 872         if (!node) {
 873                 prealloc = alloc_extent_state_atomic(prealloc);
 874                 BUG_ON(!prealloc);
 875                 err = insert_state(tree, prealloc, start, end,
 876                                    &p, &parent, &bits);
 877                 if (err)
 878                         extent_io_tree_panic(tree, err);
 879
 880                 cache_state(prealloc, cached_state);
 881                 prealloc = NULL;
 882                 goto out;
 883         }
 884         state = rb_entry(node, struct extent_state, rb_node);
 885 hit_next:
 886         last_start = state->start;
 887         last_end = state->end;
 888
 889         /*
 890          * | ---- desired range ---- |
 891          * | state |
 892          *
 893          * Just lock what we found and keep going
 894          */
 895         if (state->start == start && state->end <= end) {
 896                 if (state->state & exclusive_bits) {
 897                         *failed_start = state->start;
 898                         err = -EEXIST;
 899                         goto out;
 900                 }
 901
 902                 set_state_bits(tree, state, &bits);
 903                 cache_state(state, cached_state);
 904                 merge_state(tree, state);
 905                 if (last_end == (u64)-1)
 906                         goto out;
 907                 start = last_end + 1;
 908                 state = next_state(state);
 909                 if (start < end && state && state->start == start &&
 910                     !need_resched())
 911                         goto hit_next;
 912                 goto search_again;
 913         }
 914
 915         /*
 916          *     | ---- desired range ---- |
 917          * | state |
 918          *   or
 919          * | ------------- state -------------- |
 920          *
 921          * We need to split the extent we found, and may flip bits on
 922          * second half.
 923          *
 924          * If the extent we found extends past our
 925          * range, we just split and search again.  It'll get split
 926          * again the next time though.
 927          *
 928          * If the extent we found is inside our range, we set the
 929          * desired bit on it.
 930          */
 931         if (state->start < start) {
 932                 if (state->state & exclusive_bits) {
 933                         *failed_start = start;
 934                         err = -EEXIST;
 935                         goto out;
 936                 }
 937
 938                 prealloc = alloc_extent_state_atomic(prealloc);
 939                 BUG_ON(!prealloc);
 940                 err = split_state(tree, state, prealloc, start);
 941                 if (err)
 942                         extent_io_tree_panic(tree, err);
 943
 944                 prealloc = NULL;
 945                 if (err)
 946                         goto out;
 947                 if (state->end <= end) {
 948                         set_state_bits(tree, state, &bits);
 949                         cache_state(state, cached_state);
 950                         merge_state(tree, state);
 951                         if (last_end == (u64)-1)
 952                                 goto out;
 953                         start = last_end + 1;
 954                         state = next_state(state);
 955                         if (start < end && state && state->start == start &&
 956                             !need_resched())
 957                                 goto hit_next;
 958                 }
 959                 goto search_again;
 960         }
 961         /*
 962          * | ---- desired range ---- |
 963          *     | state | or               | state |
 964          *
 965          * There's a hole, we need to insert something in it and
 966          * ignore the extent we found.
 967          */
 968         if (state->start > start) {
 969                 u64 this_end;
 970                 if (end < last_start)
 971                         this_end = end;
 972                 else
 973                         this_end = last_start - 1;
 974
 975                 prealloc = alloc_extent_state_atomic(prealloc);
 976                 BUG_ON(!prealloc);
 977
 978                 /*
 979                  * Avoid to free 'prealloc' if it can be merged with
 980                  * the later extent.
 981                  */
 982                 err = insert_state(tree, prealloc, start, this_end,
 983                                    NULL, NULL, &bits);
 984                 if (err)
 985                         extent_io_tree_panic(tree, err);
 986
 987                 cache_state(prealloc, cached_state);
 988                 prealloc = NULL;
 989                 start = this_end + 1;
 990                 goto search_again;
 991         }
 992         /*
 993          * | ---- desired range ---- |
 994          *                        | state |
 995          * We need to split the extent, and set the bit
 996          * on the first half
 997          */
 998         if (state->start <= end && state->end > end) {
 999                 if (state->state & exclusive_bits) {
1000                         *failed_start = start;
1001                         err = -EEXIST;
1002                         goto out;
1003                 }
1004
1005                 prealloc = alloc_extent_state_atomic(prealloc);
1006                 BUG_ON(!prealloc);
1007                 err = split_state(tree, state, prealloc, end + 1);
1008                 if (err)
1009                         extent_io_tree_panic(tree, err);
1010
1011                 set_state_bits(tree, prealloc, &bits);
1012                 cache_state(prealloc, cached_state);
1013                 merge_state(tree, prealloc);
1014                 prealloc = NULL;
1015                 goto out;
1016         }
1017
1018         goto search_again;
1019
1020 out:
1021         spin_unlock(&tree->lock);
1022         if (prealloc)
1023                 free_extent_state(prealloc);
1024
1025         return err;
1026
1027 search_again:
1028         if (start > end)
1029                 goto out;
1030         spin_unlock(&tree->lock);
1031         if (mask & __GFP_WAIT)
1032                 cond_resched();
1033         goto again;
1034 }
1035
1036 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1037                    unsigned bits, u64 * failed_start,
1038                    struct extent_state **cached_state, gfp_t mask)
1039 {
1040         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1041                                 cached_state, mask);
1042 }
1043
1044
1045 /**
1046  * convert_extent_bit - convert all bits in a given range from one bit to
1047  *                      another
1048  * @tree:       the io tree to search
1049  * @start:      the start offset in bytes
1050  * @end:        the end offset in bytes (inclusive)
1051  * @bits:       the bits to set in this range
1052  * @clear_bits: the bits to clear in this range
1053  * @cached_state:       state that we're going to cache
1054  * @mask:       the allocation mask
1055  *
1056  * This will go through and set bits for the given range.  If any states exist
1057  * already in this range they are set with the given bit and cleared of the
1058  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1059  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1060  * boundary bits like LOCK.
1061  */
1062 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1063                        unsigned bits, unsigned clear_bits,
1064                        struct extent_state **cached_state, gfp_t mask)
1065 {
1066         struct extent_state *state;
1067         struct extent_state *prealloc = NULL;
1068         struct rb_node *node;
1069         struct rb_node **p;
1070         struct rb_node *parent;
1071         int err = 0;
1072         u64 last_start;
1073         u64 last_end;
1074         bool first_iteration = true;
1075
1076         btrfs_debug_check_extent_io_range(tree, start, end);
1077
1078 again:
1079         if (!prealloc && (mask & __GFP_WAIT)) {
1080                 /*
1081                  * Best effort, don't worry if extent state allocation fails
1082                  * here for the first iteration. We might have a cached state
1083                  * that matches exactly the target range, in which case no
1084                  * extent state allocations are needed. We'll only know this
1085                  * after locking the tree.
1086                  */
1087                 prealloc = alloc_extent_state(mask);
1088                 if (!prealloc && !first_iteration)
1089                         return -ENOMEM;
1090         }
1091
1092         spin_lock(&tree->lock);
1093         if (cached_state && *cached_state) {
1094                 state = *cached_state;
1095                 if (state->start <= start && state->end > start &&
1096                     extent_state_in_tree(state)) {
1097                         node = &state->rb_node;
1098                         goto hit_next;
1099                 }
1100         }
1101
1102         /*
1103          * this search will find all the extents that end after
1104          * our range starts.
1105          */
1106         node = tree_search_for_insert(tree, start, &p, &parent);
1107         if (!node) {
1108                 prealloc = alloc_extent_state_atomic(prealloc);
1109                 if (!prealloc) {
1110                         err = -ENOMEM;
1111                         goto out;
1112                 }
1113                 err = insert_state(tree, prealloc, start, end,
1114                                    &p, &parent, &bits);
1115                 if (err)
1116                         extent_io_tree_panic(tree, err);
1117                 cache_state(prealloc, cached_state);
1118                 prealloc = NULL;
1119                 goto out;
1120         }
1121         state = rb_entry(node, struct extent_state, rb_node);
1122 hit_next:
1123         last_start = state->start;
1124         last_end = state->end;
1125
1126         /*
1127          * | ---- desired range ---- |
1128          * | state |
1129          *
1130          * Just lock what we found and keep going
1131          */
1132         if (state->start == start && state->end <= end) {
1133                 set_state_bits(tree, state, &bits);
1134                 cache_state(state, cached_state);
1135                 state = clear_state_bit(tree, state, &clear_bits, 0);
1136                 if (last_end == (u64)-1)
1137                         goto out;
1138                 start = last_end + 1;
1139                 if (start < end && state && state->start == start &&
1140                     !need_resched())
1141                         goto hit_next;
1142                 goto search_again;
1143         }
1144
1145         /*
1146          *     | ---- desired range ---- |
1147          * | state |
1148          *   or
1149          * | ------------- state -------------- |
1150          *
1151          * We need to split the extent we found, and may flip bits on
1152          * second half.
1153          *
1154          * If the extent we found extends past our
1155          * range, we just split and search again.  It'll get split
1156          * again the next time though.
1157          *
1158          * If the extent we found is inside our range, we set the
1159          * desired bit on it.
1160          */
1161         if (state->start < start) {
1162                 prealloc = alloc_extent_state_atomic(prealloc);
1163                 if (!prealloc) {
1164                         err = -ENOMEM;
1165                         goto out;
1166                 }
1167                 err = split_state(tree, state, prealloc, start);
1168                 if (err)
1169                         extent_io_tree_panic(tree, err);
1170                 prealloc = NULL;
1171                 if (err)
1172                         goto out;
1173                 if (state->end <= end) {
1174                         set_state_bits(tree, state, &bits);
1175                         cache_state(state, cached_state);
1176                         state = clear_state_bit(tree, state, &clear_bits, 0);
1177                         if (last_end == (u64)-1)
1178                                 goto out;
1179                         start = last_end + 1;
1180                         if (start < end && state && state->start == start &&
1181                             !need_resched())
1182                                 goto hit_next;
1183                 }
1184                 goto search_again;
1185         }
1186         /*
1187          * | ---- desired range ---- |
1188          *     | state | or               | state |
1189          *
1190          * There's a hole, we need to insert something in it and
1191          * ignore the extent we found.
1192          */
1193         if (state->start > start) {
1194                 u64 this_end;
1195                 if (end < last_start)
1196                         this_end = end;
1197                 else
1198                         this_end = last_start - 1;
1199
1200                 prealloc = alloc_extent_state_atomic(prealloc);
1201                 if (!prealloc) {
1202                         err = -ENOMEM;
1203                         goto out;
1204                 }
1205
1206                 /*
1207                  * Avoid to free 'prealloc' if it can be merged with
1208                  * the later extent.
1209                  */
1210                 err = insert_state(tree, prealloc, start, this_end,
1211                                    NULL, NULL, &bits);
1212                 if (err)
1213                         extent_io_tree_panic(tree, err);
1214                 cache_state(prealloc, cached_state);
1215                 prealloc = NULL;
1216                 start = this_end + 1;
1217                 goto search_again;
1218         }
1219         /*
1220          * | ---- desired range ---- |
1221          *                        | state |
1222          * We need to split the extent, and set the bit
1223          * on the first half
1224          */
1225         if (state->start <= end && state->end > end) {
1226                 prealloc = alloc_extent_state_atomic(prealloc);
1227                 if (!prealloc) {
1228                         err = -ENOMEM;
1229                         goto out;
1230                 }
1231
1232                 err = split_state(tree, state, prealloc, end + 1);
1233                 if (err)
1234                         extent_io_tree_panic(tree, err);
1235
1236                 set_state_bits(tree, prealloc, &bits);
1237                 cache_state(prealloc, cached_state);
1238                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1239                 prealloc = NULL;
1240                 goto out;
1241         }
1242
1243         goto search_again;
1244
1245 out:
1246         spin_unlock(&tree->lock);
1247         if (prealloc)
1248                 free_extent_state(prealloc);
1249
1250         return err;
1251
1252 search_again:
1253         if (start > end)
1254                 goto out;
1255         spin_unlock(&tree->lock);
1256         if (mask & __GFP_WAIT)
1257                 cond_resched();
1258         first_iteration = false;
1259         goto again;
1260 }
1261
1262 /* wrappers around set/clear extent bit */
1263 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1264                      gfp_t mask)
1265 {
1266         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1267                               NULL, mask);
1268 }
1269
1270 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1271                     unsigned bits, gfp_t mask)
1272 {
1273         return set_extent_bit(tree, start, end, bits, NULL,
1274                               NULL, mask);
1275 }
1276
1277 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1278                       unsigned bits, gfp_t mask)
1279 {
1280         int wake = 0;
1281
1282         if (bits & EXTENT_LOCKED)
1283                 wake = 1;
1284
1285         return clear_extent_bit(tree, start, end, bits, wake, 0, NULL, mask);
1286 }
1287
1288 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1289                         struct extent_state **cached_state, gfp_t mask)
1290 {
1291         return set_extent_bit(tree, start, end,
1292                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1293                               NULL, cached_state, mask);
1294 }
1295
1296 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1297                       struct extent_state **cached_state, gfp_t mask)
1298 {
1299         return set_extent_bit(tree, start, end,
1300                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1301                               NULL, cached_state, mask);
1302 }
1303
1304 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1305                        gfp_t mask)
1306 {
1307         return clear_extent_bit(tree, start, end,
1308                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1309                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1310 }
1311
1312 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1313                      gfp_t mask)
1314 {
1315         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1316                               NULL, mask);
1317 }
1318
1319 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1320                         struct extent_state **cached_state, gfp_t mask)
1321 {
1322         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1323                               cached_state, mask);
1324 }
1325
1326 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1327                           struct extent_state **cached_state, gfp_t mask)
1328 {
1329         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1330                                 cached_state, mask);
1331 }
1332
1333 /*
1334  * either insert or lock state struct between start and end use mask to tell
1335  * us if waiting is desired.
1336  */
1337 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1338                      unsigned bits, struct extent_state **cached_state)
1339 {
1340         int err;
1341         u64 failed_start;
1342
1343         while (1) {
1344                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1345                                        EXTENT_LOCKED, &failed_start,
1346                                        cached_state, GFP_NOFS);
1347                 if (err == -EEXIST) {
1348                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1349                         start = failed_start;
1350                 } else
1351                         break;
1352                 WARN_ON(start > end);
1353         }
1354         return err;
1355 }
1356
1357 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1358 {
1359         return lock_extent_bits(tree, start, end, 0, NULL);
1360 }
1361
1362 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1363 {
1364         int err;
1365         u64 failed_start;
1366
1367         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1368                                &failed_start, NULL, GFP_NOFS);
1369         if (err == -EEXIST) {
1370                 if (failed_start > start)
1371                         clear_extent_bit(tree, start, failed_start - 1,
1372                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1373                 return 0;
1374         }
1375         return 1;
1376 }
1377
1378 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1379                          struct extent_state **cached, gfp_t mask)
1380 {
1381         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1382                                 mask);
1383 }
1384
1385 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1386 {
1387         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1388                                 GFP_NOFS);
1389 }
1390
1391 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1392 {
1393         unsigned long index = start >> PAGE_CACHE_SHIFT;
1394         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1395         struct page *page;
1396
1397         while (index <= end_index) {
1398                 page = find_get_page(inode->i_mapping, index);
1399                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1400                 clear_page_dirty_for_io(page);
1401                 page_cache_release(page);
1402                 index++;
1403         }
1404         return 0;
1405 }
1406
1407 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1408 {
1409         unsigned long index = start >> PAGE_CACHE_SHIFT;
1410         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1411         struct page *page;
1412
1413         while (index <= end_index) {
1414                 page = find_get_page(inode->i_mapping, index);
1415                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1416                 __set_page_dirty_nobuffers(page);
1417                 account_page_redirty(page);
1418                 page_cache_release(page);
1419                 index++;
1420         }
1421         return 0;
1422 }
1423
1424 /*
1425  * helper function to set both pages and extents in the tree writeback
1426  */
1427 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1428 {
1429         unsigned long index = start >> PAGE_CACHE_SHIFT;
1430         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1431         struct page *page;
1432
1433         while (index <= end_index) {
1434                 page = find_get_page(tree->mapping, index);
1435                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1436                 set_page_writeback(page);
1437                 page_cache_release(page);
1438                 index++;
1439         }
1440         return 0;
1441 }
1442
1443 /* find the first state struct with 'bits' set after 'start', and
1444  * return it.  tree->lock must be held.  NULL will returned if
1445  * nothing was found after 'start'
1446  */
1447 static struct extent_state *
1448 find_first_extent_bit_state(struct extent_io_tree *tree,
1449                             u64 start, unsigned bits)
1450 {
1451         struct rb_node *node;
1452         struct extent_state *state;
1453
1454         /*
1455          * this search will find all the extents that end after
1456          * our range starts.
1457          */
1458         node = tree_search(tree, start);
1459         if (!node)
1460                 goto out;
1461
1462         while (1) {
1463                 state = rb_entry(node, struct extent_state, rb_node);
1464                 if (state->end >= start && (state->state & bits))
1465                         return state;
1466
1467                 node = rb_next(node);
1468                 if (!node)
1469                         break;
1470         }
1471 out:
1472         return NULL;
1473 }
1474
1475 /*
1476  * find the first offset in the io tree with 'bits' set. zero is
1477  * returned if we find something, and *start_ret and *end_ret are
1478  * set to reflect the state struct that was found.
1479  *
1480  * If nothing was found, 1 is returned. If found something, return 0.
1481  */
1482 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1483                           u64 *start_ret, u64 *end_ret, unsigned bits,
1484                           struct extent_state **cached_state)
1485 {
1486         struct extent_state *state;
1487         struct rb_node *n;
1488         int ret = 1;
1489
1490         spin_lock(&tree->lock);
1491         if (cached_state && *cached_state) {
1492                 state = *cached_state;
1493                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1494                         n = rb_next(&state->rb_node);
1495                         while (n) {
1496                                 state = rb_entry(n, struct extent_state,
1497                                                  rb_node);
1498                                 if (state->state & bits)
1499                                         goto got_it;
1500                                 n = rb_next(n);
1501                         }
1502                         free_extent_state(*cached_state);
1503                         *cached_state = NULL;
1504                         goto out;
1505                 }
1506                 free_extent_state(*cached_state);
1507                 *cached_state = NULL;
1508         }
1509
1510         state = find_first_extent_bit_state(tree, start, bits);
1511 got_it:
1512         if (state) {
1513                 cache_state_if_flags(state, cached_state, 0);
1514                 *start_ret = state->start;
1515                 *end_ret = state->end;
1516                 ret = 0;
1517         }
1518 out:
1519         spin_unlock(&tree->lock);
1520         return ret;
1521 }
1522
1523 /*
1524  * find a contiguous range of bytes in the file marked as delalloc, not
1525  * more than 'max_bytes'.  start and end are used to return the range,
1526  *
1527  * 1 is returned if we find something, 0 if nothing was in the tree
1528  */
1529 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1530                                         u64 *start, u64 *end, u64 max_bytes,
1531                                         struct extent_state **cached_state)
1532 {
1533         struct rb_node *node;
1534         struct extent_state *state;
1535         u64 cur_start = *start;
1536         u64 found = 0;
1537         u64 total_bytes = 0;
1538
1539         spin_lock(&tree->lock);
1540
1541         /*
1542          * this search will find all the extents that end after
1543          * our range starts.
1544          */
1545         node = tree_search(tree, cur_start);
1546         if (!node) {
1547                 if (!found)
1548                         *end = (u64)-1;
1549                 goto out;
1550         }
1551
1552         while (1) {
1553                 state = rb_entry(node, struct extent_state, rb_node);
1554                 if (found && (state->start != cur_start ||
1555                               (state->state & EXTENT_BOUNDARY))) {
1556                         goto out;
1557                 }
1558                 if (!(state->state & EXTENT_DELALLOC)) {
1559                         if (!found)
1560                                 *end = state->end;
1561                         goto out;
1562                 }
1563                 if (!found) {
1564                         *start = state->start;
1565                         *cached_state = state;
1566                         atomic_inc(&state->refs);
1567                 }
1568                 found++;
1569                 *end = state->end;
1570                 cur_start = state->end + 1;
1571                 node = rb_next(node);
1572                 total_bytes += state->end - state->start + 1;
1573                 if (total_bytes >= max_bytes)
1574                         break;
1575                 if (!node)
1576                         break;
1577         }
1578 out:
1579         spin_unlock(&tree->lock);
1580         return found;
1581 }
1582
1583 static noinline void __unlock_for_delalloc(struct inode *inode,
1584                                            struct page *locked_page,
1585                                            u64 start, u64 end)
1586 {
1587         int ret;
1588         struct page *pages[16];
1589         unsigned long index = start >> PAGE_CACHE_SHIFT;
1590         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1591         unsigned long nr_pages = end_index - index + 1;
1592         int i;
1593
1594         if (index == locked_page->index && end_index == index)
1595                 return;
1596
1597         while (nr_pages > 0) {
1598                 ret = find_get_pages_contig(inode->i_mapping, index,
1599                                      min_t(unsigned long, nr_pages,
1600                                      ARRAY_SIZE(pages)), pages);
1601                 for (i = 0; i < ret; i++) {
1602                         if (pages[i] != locked_page)
1603                                 unlock_page(pages[i]);
1604                         page_cache_release(pages[i]);
1605                 }
1606                 nr_pages -= ret;
1607                 index += ret;
1608                 cond_resched();
1609         }
1610 }
1611
1612 static noinline int lock_delalloc_pages(struct inode *inode,
1613                                         struct page *locked_page,
1614                                         u64 delalloc_start,
1615                                         u64 delalloc_end)
1616 {
1617         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1618         unsigned long start_index = index;
1619         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1620         unsigned long pages_locked = 0;
1621         struct page *pages[16];
1622         unsigned long nrpages;
1623         int ret;
1624         int i;
1625
1626         /* the caller is responsible for locking the start index */
1627         if (index == locked_page->index && index == end_index)
1628                 return 0;
1629
1630         /* skip the page at the start index */
1631         nrpages = end_index - index + 1;
1632         while (nrpages > 0) {
1633                 ret = find_get_pages_contig(inode->i_mapping, index,
1634                                      min_t(unsigned long,
1635                                      nrpages, ARRAY_SIZE(pages)), pages);
1636                 if (ret == 0) {
1637                         ret = -EAGAIN;
1638                         goto done;
1639                 }
1640                 /* now we have an array of pages, lock them all */
1641                 for (i = 0; i < ret; i++) {
1642                         /*
1643                          * the caller is taking responsibility for
1644                          * locked_page
1645                          */
1646                         if (pages[i] != locked_page) {
1647                                 lock_page(pages[i]);
1648                                 if (!PageDirty(pages[i]) ||
1649                                     pages[i]->mapping != inode->i_mapping) {
1650                                         ret = -EAGAIN;
1651                                         unlock_page(pages[i]);
1652                                         page_cache_release(pages[i]);
1653                                         goto done;
1654                                 }
1655                         }
1656                         page_cache_release(pages[i]);
1657                         pages_locked++;
1658                 }
1659                 nrpages -= ret;
1660                 index += ret;
1661                 cond_resched();
1662         }
1663         ret = 0;
1664 done:
1665         if (ret && pages_locked) {
1666                 __unlock_for_delalloc(inode, locked_page,
1667                               delalloc_start,
1668                               ((u64)(start_index + pages_locked - 1)) <<
1669                               PAGE_CACHE_SHIFT);
1670         }
1671         return ret;
1672 }
1673
1674 /*
1675  * find a contiguous range of bytes in the file marked as delalloc, not
1676  * more than 'max_bytes'.  start and end are used to return the range,
1677  *
1678  * 1 is returned if we find something, 0 if nothing was in the tree
1679  */
1680 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1681                                     struct extent_io_tree *tree,
1682                                     struct page *locked_page, u64 *start,
1683                                     u64 *end, u64 max_bytes)
1684 {
1685         u64 delalloc_start;
1686         u64 delalloc_end;
1687         u64 found;
1688         struct extent_state *cached_state = NULL;
1689         int ret;
1690         int loops = 0;
1691
1692 again:
1693         /* step one, find a bunch of delalloc bytes starting at start */
1694         delalloc_start = *start;
1695         delalloc_end = 0;
1696         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1697                                     max_bytes, &cached_state);
1698         if (!found || delalloc_end <= *start) {
1699                 *start = delalloc_start;
1700                 *end = delalloc_end;
1701                 free_extent_state(cached_state);
1702                 return 0;
1703         }
1704
1705         /*
1706          * start comes from the offset of locked_page.  We have to lock
1707          * pages in order, so we can't process delalloc bytes before
1708          * locked_page
1709          */
1710         if (delalloc_start < *start)
1711                 delalloc_start = *start;
1712
1713         /*
1714          * make sure to limit the number of pages we try to lock down
1715          */
1716         if (delalloc_end + 1 - delalloc_start > max_bytes)
1717                 delalloc_end = delalloc_start + max_bytes - 1;
1718
1719         /* step two, lock all the pages after the page that has start */
1720         ret = lock_delalloc_pages(inode, locked_page,
1721                                   delalloc_start, delalloc_end);
1722         if (ret == -EAGAIN) {
1723                 /* some of the pages are gone, lets avoid looping by
1724                  * shortening the size of the delalloc range we're searching
1725                  */
1726                 free_extent_state(cached_state);
1727                 cached_state = NULL;
1728                 if (!loops) {
1729                         max_bytes = PAGE_CACHE_SIZE;
1730                         loops = 1;
1731                         goto again;
1732                 } else {
1733                         found = 0;
1734                         goto out_failed;
1735                 }
1736         }
1737         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1738
1739         /* step three, lock the state bits for the whole range */
1740         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1741
1742         /* then test to make sure it is all still delalloc */
1743         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1744                              EXTENT_DELALLOC, 1, cached_state);
1745         if (!ret) {
1746                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1747                                      &cached_state, GFP_NOFS);
1748                 __unlock_for_delalloc(inode, locked_page,
1749                               delalloc_start, delalloc_end);
1750                 cond_resched();
1751                 goto again;
1752         }
1753         free_extent_state(cached_state);
1754         *start = delalloc_start;
1755         *end = delalloc_end;
1756 out_failed:
1757         return found;
1758 }
1759
1760 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1761                                  struct page *locked_page,
1762                                  unsigned clear_bits,
1763                                  unsigned long page_ops)
1764 {
1765         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1766         int ret;
1767         struct page *pages[16];
1768         unsigned long index = start >> PAGE_CACHE_SHIFT;
1769         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1770         unsigned long nr_pages = end_index - index + 1;
1771         int i;
1772
1773         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1774         if (page_ops == 0)
1775                 return 0;
1776
1777         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1778                 mapping_set_error(inode->i_mapping, -EIO);
1779
1780         while (nr_pages > 0) {
1781                 ret = find_get_pages_contig(inode->i_mapping, index,
1782                                      min_t(unsigned long,
1783                                      nr_pages, ARRAY_SIZE(pages)), pages);
1784                 for (i = 0; i < ret; i++) {
1785
1786                         if (page_ops & PAGE_SET_PRIVATE2)
1787                                 SetPagePrivate2(pages[i]);
1788
1789                         if (pages[i] == locked_page) {
1790                                 page_cache_release(pages[i]);
1791                                 continue;
1792                         }
1793                         if (page_ops & PAGE_CLEAR_DIRTY)
1794                                 clear_page_dirty_for_io(pages[i]);
1795                         if (page_ops & PAGE_SET_WRITEBACK)
1796                                 set_page_writeback(pages[i]);
1797                         if (page_ops & PAGE_SET_ERROR)
1798                                 SetPageError(pages[i]);
1799                         if (page_ops & PAGE_END_WRITEBACK)
1800                                 end_page_writeback(pages[i]);
1801                         if (page_ops & PAGE_UNLOCK)
1802                                 unlock_page(pages[i]);
1803                         page_cache_release(pages[i]);
1804                 }
1805                 nr_pages -= ret;
1806                 index += ret;
1807                 cond_resched();
1808         }
1809         return 0;
1810 }
1811
1812 /*
1813  * count the number of bytes in the tree that have a given bit(s)
1814  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1815  * cached.  The total number found is returned.
1816  */
1817 u64 count_range_bits(struct extent_io_tree *tree,
1818                      u64 *start, u64 search_end, u64 max_bytes,
1819                      unsigned bits, int contig)
1820 {
1821         struct rb_node *node;
1822         struct extent_state *state;
1823         u64 cur_start = *start;
1824         u64 total_bytes = 0;
1825         u64 last = 0;
1826         int found = 0;
1827
1828         if (WARN_ON(search_end <= cur_start))
1829                 return 0;
1830
1831         spin_lock(&tree->lock);
1832         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1833                 total_bytes = tree->dirty_bytes;
1834                 goto out;
1835         }
1836         /*
1837          * this search will find all the extents that end after
1838          * our range starts.
1839          */
1840         node = tree_search(tree, cur_start);
1841         if (!node)
1842                 goto out;
1843
1844         while (1) {
1845                 state = rb_entry(node, struct extent_state, rb_node);
1846                 if (state->start > search_end)
1847                         break;
1848                 if (contig && found && state->start > last + 1)
1849                         break;
1850                 if (state->end >= cur_start && (state->state & bits) == bits) {
1851                         total_bytes += min(search_end, state->end) + 1 -
1852                                        max(cur_start, state->start);
1853                         if (total_bytes >= max_bytes)
1854                                 break;
1855                         if (!found) {
1856                                 *start = max(cur_start, state->start);
1857                                 found = 1;
1858                         }
1859                         last = state->end;
1860                 } else if (contig && found) {
1861                         break;
1862                 }
1863                 node = rb_next(node);
1864                 if (!node)
1865                         break;
1866         }
1867 out:
1868         spin_unlock(&tree->lock);
1869         return total_bytes;
1870 }
1871
1872 /*
1873  * set the private field for a given byte offset in the tree.  If there isn't
1874  * an extent_state there already, this does nothing.
1875  */
1876 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1877 {
1878         struct rb_node *node;
1879         struct extent_state *state;
1880         int ret = 0;
1881
1882         spin_lock(&tree->lock);
1883         /*
1884          * this search will find all the extents that end after
1885          * our range starts.
1886          */
1887         node = tree_search(tree, start);
1888         if (!node) {
1889                 ret = -ENOENT;
1890                 goto out;
1891         }
1892         state = rb_entry(node, struct extent_state, rb_node);
1893         if (state->start != start) {
1894                 ret = -ENOENT;
1895                 goto out;
1896         }
1897         state->private = private;
1898 out:
1899         spin_unlock(&tree->lock);
1900         return ret;
1901 }
1902
1903 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1904 {
1905         struct rb_node *node;
1906         struct extent_state *state;
1907         int ret = 0;
1908
1909         spin_lock(&tree->lock);
1910         /*
1911          * this search will find all the extents that end after
1912          * our range starts.
1913          */
1914         node = tree_search(tree, start);
1915         if (!node) {
1916                 ret = -ENOENT;
1917                 goto out;
1918         }
1919         state = rb_entry(node, struct extent_state, rb_node);
1920         if (state->start != start) {
1921                 ret = -ENOENT;
1922                 goto out;
1923         }
1924         *private = state->private;
1925 out:
1926         spin_unlock(&tree->lock);
1927         return ret;
1928 }
1929
1930 /*
1931  * searches a range in the state tree for a given mask.
1932  * If 'filled' == 1, this returns 1 only if every extent in the tree
1933  * has the bits set.  Otherwise, 1 is returned if any bit in the
1934  * range is found set.
1935  */
1936 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1937                    unsigned bits, int filled, struct extent_state *cached)
1938 {
1939         struct extent_state *state = NULL;
1940         struct rb_node *node;
1941         int bitset = 0;
1942
1943         spin_lock(&tree->lock);
1944         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1945             cached->end > start)
1946                 node = &cached->rb_node;
1947         else
1948                 node = tree_search(tree, start);
1949         while (node && start <= end) {
1950                 state = rb_entry(node, struct extent_state, rb_node);
1951
1952                 if (filled && state->start > start) {
1953                         bitset = 0;
1954                         break;
1955                 }
1956
1957                 if (state->start > end)
1958                         break;
1959
1960                 if (state->state & bits) {
1961                         bitset = 1;
1962                         if (!filled)
1963                                 break;
1964                 } else if (filled) {
1965                         bitset = 0;
1966                         break;
1967                 }
1968
1969                 if (state->end == (u64)-1)
1970                         break;
1971
1972                 start = state->end + 1;
1973                 if (start > end)
1974                         break;
1975                 node = rb_next(node);
1976                 if (!node) {
1977                         if (filled)
1978                                 bitset = 0;
1979                         break;
1980                 }
1981         }
1982         spin_unlock(&tree->lock);
1983         return bitset;
1984 }
1985
1986 /*
1987  * helper function to set a given page up to date if all the
1988  * extents in the tree for that page are up to date
1989  */
1990 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1991 {
1992         u64 start = page_offset(page);
1993         u64 end = start + PAGE_CACHE_SIZE - 1;
1994         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1995                 SetPageUptodate(page);
1996 }
1997
1998 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1999 {
2000         int ret;
2001         int err = 0;
2002         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2003
2004         set_state_private(failure_tree, rec->start, 0);
2005         ret = clear_extent_bits(failure_tree, rec->start,
2006                                 rec->start + rec->len - 1,
2007                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2008         if (ret)
2009                 err = ret;
2010
2011         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
2012                                 rec->start + rec->len - 1,
2013                                 EXTENT_DAMAGED, GFP_NOFS);
2014         if (ret && !err)
2015                 err = ret;
2016
2017         kfree(rec);
2018         return err;
2019 }
2020
2021 /*
2022  * this bypasses the standard btrfs submit functions deliberately, as
2023  * the standard behavior is to write all copies in a raid setup. here we only
2024  * want to write the one bad copy. so we do the mapping for ourselves and issue
2025  * submit_bio directly.
2026  * to avoid any synchronization issues, wait for the data after writing, which
2027  * actually prevents the read that triggered the error from finishing.
2028  * currently, there can be no more than two copies of every data bit. thus,
2029  * exactly one rewrite is required.
2030  */
2031 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2032                       struct page *page, unsigned int pg_offset, int mirror_num)
2033 {
2034         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2035         struct bio *bio;
2036         struct btrfs_device *dev;
2037         u64 map_length = 0;
2038         u64 sector;
2039         struct btrfs_bio *bbio = NULL;
2040         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2041         int ret;
2042
2043         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2044         BUG_ON(!mirror_num);
2045
2046         /* we can't repair anything in raid56 yet */
2047         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2048                 return 0;
2049
2050         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2051         if (!bio)
2052                 return -EIO;
2053         bio->bi_iter.bi_size = 0;
2054         map_length = length;
2055
2056         ret = btrfs_map_block(fs_info, WRITE, logical,
2057                               &map_length, &bbio, mirror_num);
2058         if (ret) {
2059                 bio_put(bio);
2060                 return -EIO;
2061         }
2062         BUG_ON(mirror_num != bbio->mirror_num);
2063         sector = bbio->stripes[mirror_num-1].physical >> 9;
2064         bio->bi_iter.bi_sector = sector;
2065         dev = bbio->stripes[mirror_num-1].dev;
2066         btrfs_put_bbio(bbio);
2067         if (!dev || !dev->bdev || !dev->writeable) {
2068                 bio_put(bio);
2069                 return -EIO;
2070         }
2071         bio->bi_bdev = dev->bdev;
2072         bio_add_page(bio, page, length, pg_offset);
2073
2074         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio)) {
2075                 /* try to remap that extent elsewhere? */
2076                 bio_put(bio);
2077                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2078                 return -EIO;
2079         }
2080
2081         printk_ratelimited_in_rcu(KERN_INFO
2082                                   "BTRFS: read error corrected: ino %llu off %llu (dev %s sector %llu)\n",
2083                                   btrfs_ino(inode), start,
2084                                   rcu_str_deref(dev->name), sector);
2085         bio_put(bio);
2086         return 0;
2087 }
2088
2089 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2090                          int mirror_num)
2091 {
2092         u64 start = eb->start;
2093         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2094         int ret = 0;
2095
2096         if (root->fs_info->sb->s_flags & MS_RDONLY)
2097                 return -EROFS;
2098
2099         for (i = 0; i < num_pages; i++) {
2100                 struct page *p = eb->pages[i];
2101
2102                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2103                                         PAGE_CACHE_SIZE, start, p,
2104                                         start - page_offset(p), mirror_num);
2105                 if (ret)
2106                         break;
2107                 start += PAGE_CACHE_SIZE;
2108         }
2109
2110         return ret;
2111 }
2112
2113 /*
2114  * each time an IO finishes, we do a fast check in the IO failure tree
2115  * to see if we need to process or clean up an io_failure_record
2116  */
2117 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2118                      unsigned int pg_offset)
2119 {
2120         u64 private;
2121         u64 private_failure;
2122         struct io_failure_record *failrec;
2123         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2124         struct extent_state *state;
2125         int num_copies;
2126         int ret;
2127
2128         private = 0;
2129         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2130                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2131         if (!ret)
2132                 return 0;
2133
2134         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2135                                 &private_failure);
2136         if (ret)
2137                 return 0;
2138
2139         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2140         BUG_ON(!failrec->this_mirror);
2141
2142         if (failrec->in_validation) {
2143                 /* there was no real error, just free the record */
2144                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2145                          failrec->start);
2146                 goto out;
2147         }
2148         if (fs_info->sb->s_flags & MS_RDONLY)
2149                 goto out;
2150
2151         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2152         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2153                                             failrec->start,
2154                                             EXTENT_LOCKED);
2155         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2156
2157         if (state && state->start <= failrec->start &&
2158             state->end >= failrec->start + failrec->len - 1) {
2159                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2160                                               failrec->len);
2161                 if (num_copies > 1)  {
2162                         repair_io_failure(inode, start, failrec->len,
2163                                           failrec->logical, page,
2164                                           pg_offset, failrec->failed_mirror);
2165                 }
2166         }
2167
2168 out:
2169         free_io_failure(inode, failrec);
2170
2171         return 0;
2172 }
2173
2174 /*
2175  * Can be called when
2176  * - hold extent lock
2177  * - under ordered extent
2178  * - the inode is freeing
2179  */
2180 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2181 {
2182         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2183         struct io_failure_record *failrec;
2184         struct extent_state *state, *next;
2185
2186         if (RB_EMPTY_ROOT(&failure_tree->state))
2187                 return;
2188
2189         spin_lock(&failure_tree->lock);
2190         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2191         while (state) {
2192                 if (state->start > end)
2193                         break;
2194
2195                 ASSERT(state->end <= end);
2196
2197                 next = next_state(state);
2198
2199                 failrec = (struct io_failure_record *)(unsigned long)state->private;
2200                 free_extent_state(state);
2201                 kfree(failrec);
2202
2203                 state = next;
2204         }
2205         spin_unlock(&failure_tree->lock);
2206 }
2207
2208 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2209                                 struct io_failure_record **failrec_ret)
2210 {
2211         struct io_failure_record *failrec;
2212         u64 private;
2213         struct extent_map *em;
2214         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2215         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2216         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2217         int ret;
2218         u64 logical;
2219
2220         ret = get_state_private(failure_tree, start, &private);
2221         if (ret) {
2222                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2223                 if (!failrec)
2224                         return -ENOMEM;
2225
2226                 failrec->start = start;
2227                 failrec->len = end - start + 1;
2228                 failrec->this_mirror = 0;
2229                 failrec->bio_flags = 0;
2230                 failrec->in_validation = 0;
2231
2232                 read_lock(&em_tree->lock);
2233                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2234                 if (!em) {
2235                         read_unlock(&em_tree->lock);
2236                         kfree(failrec);
2237                         return -EIO;
2238                 }
2239
2240                 if (em->start > start || em->start + em->len <= start) {
2241                         free_extent_map(em);
2242                         em = NULL;
2243                 }
2244                 read_unlock(&em_tree->lock);
2245                 if (!em) {
2246                         kfree(failrec);
2247                         return -EIO;
2248                 }
2249
2250                 logical = start - em->start;
2251                 logical = em->block_start + logical;
2252                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2253                         logical = em->block_start;
2254                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2255                         extent_set_compress_type(&failrec->bio_flags,
2256                                                  em->compress_type);
2257                 }
2258
2259                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2260                          logical, start, failrec->len);
2261
2262                 failrec->logical = logical;
2263                 free_extent_map(em);
2264
2265                 /* set the bits in the private failure tree */
2266                 ret = set_extent_bits(failure_tree, start, end,
2267                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2268                 if (ret >= 0)
2269                         ret = set_state_private(failure_tree, start,
2270                                                 (u64)(unsigned long)failrec);
2271                 /* set the bits in the inode's tree */
2272                 if (ret >= 0)
2273                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2274                                                 GFP_NOFS);
2275                 if (ret < 0) {
2276                         kfree(failrec);
2277                         return ret;
2278                 }
2279         } else {
2280                 failrec = (struct io_failure_record *)(unsigned long)private;
2281                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2282                          failrec->logical, failrec->start, failrec->len,
2283                          failrec->in_validation);
2284                 /*
2285                  * when data can be on disk more than twice, add to failrec here
2286                  * (e.g. with a list for failed_mirror) to make
2287                  * clean_io_failure() clean all those errors at once.
2288                  */
2289         }
2290
2291         *failrec_ret = failrec;
2292
2293         return 0;
2294 }
2295
2296 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2297                            struct io_failure_record *failrec, int failed_mirror)
2298 {
2299         int num_copies;
2300
2301         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2302                                       failrec->logical, failrec->len);
2303         if (num_copies == 1) {
2304                 /*
2305                  * we only have a single copy of the data, so don't bother with
2306                  * all the retry and error correction code that follows. no
2307                  * matter what the error is, it is very likely to persist.
2308                  */
2309                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2310                          num_copies, failrec->this_mirror, failed_mirror);
2311                 return 0;
2312         }
2313
2314         /*
2315          * there are two premises:
2316          *      a) deliver good data to the caller
2317          *      b) correct the bad sectors on disk
2318          */
2319         if (failed_bio->bi_vcnt > 1) {
2320                 /*
2321                  * to fulfill b), we need to know the exact failing sectors, as
2322                  * we don't want to rewrite any more than the failed ones. thus,
2323                  * we need separate read requests for the failed bio
2324                  *
2325                  * if the following BUG_ON triggers, our validation request got
2326                  * merged. we need separate requests for our algorithm to work.
2327                  */
2328                 BUG_ON(failrec->in_validation);
2329                 failrec->in_validation = 1;
2330                 failrec->this_mirror = failed_mirror;
2331         } else {
2332                 /*
2333                  * we're ready to fulfill a) and b) alongside. get a good copy
2334                  * of the failed sector and if we succeed, we have setup
2335                  * everything for repair_io_failure to do the rest for us.
2336                  */
2337                 if (failrec->in_validation) {
2338                         BUG_ON(failrec->this_mirror != failed_mirror);
2339                         failrec->in_validation = 0;
2340                         failrec->this_mirror = 0;
2341                 }
2342                 failrec->failed_mirror = failed_mirror;
2343                 failrec->this_mirror++;
2344                 if (failrec->this_mirror == failed_mirror)
2345                         failrec->this_mirror++;
2346         }
2347
2348         if (failrec->this_mirror > num_copies) {
2349                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2350                          num_copies, failrec->this_mirror, failed_mirror);
2351                 return 0;
2352         }
2353
2354         return 1;
2355 }
2356
2357
2358 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2359                                     struct io_failure_record *failrec,
2360                                     struct page *page, int pg_offset, int icsum,
2361                                     bio_end_io_t *endio_func, void *data)
2362 {
2363         struct bio *bio;
2364         struct btrfs_io_bio *btrfs_failed_bio;
2365         struct btrfs_io_bio *btrfs_bio;
2366
2367         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2368         if (!bio)
2369                 return NULL;
2370
2371         bio->bi_end_io = endio_func;
2372         bio->bi_iter.bi_sector = failrec->logical >> 9;
2373         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2374         bio->bi_iter.bi_size = 0;
2375         bio->bi_private = data;
2376
2377         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2378         if (btrfs_failed_bio->csum) {
2379                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2380                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2381
2382                 btrfs_bio = btrfs_io_bio(bio);
2383                 btrfs_bio->csum = btrfs_bio->csum_inline;
2384                 icsum *= csum_size;
2385                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2386                        csum_size);
2387         }
2388
2389         bio_add_page(bio, page, failrec->len, pg_offset);
2390
2391         return bio;
2392 }
2393
2394 /*
2395  * this is a generic handler for readpage errors (default
2396  * readpage_io_failed_hook). if other copies exist, read those and write back
2397  * good data to the failed position. does not investigate in remapping the
2398  * failed extent elsewhere, hoping the device will be smart enough to do this as
2399  * needed
2400  */
2401
2402 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2403                               struct page *page, u64 start, u64 end,
2404                               int failed_mirror)
2405 {
2406         struct io_failure_record *failrec;
2407         struct inode *inode = page->mapping->host;
2408         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2409         struct bio *bio;
2410         int read_mode;
2411         int ret;
2412
2413         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2414
2415         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2416         if (ret)
2417                 return ret;
2418
2419         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2420         if (!ret) {
2421                 free_io_failure(inode, failrec);
2422                 return -EIO;
2423         }
2424
2425         if (failed_bio->bi_vcnt > 1)
2426                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2427         else
2428                 read_mode = READ_SYNC;
2429
2430         phy_offset >>= inode->i_sb->s_blocksize_bits;
2431         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2432                                       start - page_offset(page),
2433                                       (int)phy_offset, failed_bio->bi_end_io,
2434                                       NULL);
2435         if (!bio) {
2436                 free_io_failure(inode, failrec);
2437                 return -EIO;
2438         }
2439
2440         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2441                  read_mode, failrec->this_mirror, failrec->in_validation);
2442
2443         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2444                                          failrec->this_mirror,
2445                                          failrec->bio_flags, 0);
2446         if (ret) {
2447                 free_io_failure(inode, failrec);
2448                 bio_put(bio);
2449         }
2450
2451         return ret;
2452 }
2453
2454 /* lots and lots of room for performance fixes in the end_bio funcs */
2455
2456 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2457 {
2458         int uptodate = (err == 0);
2459         struct extent_io_tree *tree;
2460         int ret = 0;
2461
2462         tree = &BTRFS_I(page->mapping->host)->io_tree;
2463
2464         if (tree->ops && tree->ops->writepage_end_io_hook) {
2465                 ret = tree->ops->writepage_end_io_hook(page, start,
2466                                                end, NULL, uptodate);
2467                 if (ret)
2468                         uptodate = 0;
2469         }
2470
2471         if (!uptodate) {
2472                 ClearPageUptodate(page);
2473                 SetPageError(page);
2474                 ret = ret < 0 ? ret : -EIO;
2475                 mapping_set_error(page->mapping, ret);
2476         }
2477         return 0;
2478 }
2479
2480 /*
2481  * after a writepage IO is done, we need to:
2482  * clear the uptodate bits on error
2483  * clear the writeback bits in the extent tree for this IO
2484  * end_page_writeback if the page has no more pending IO
2485  *
2486  * Scheduling is not allowed, so the extent state tree is expected
2487  * to have one and only one object corresponding to this IO.
2488  */
2489 static void end_bio_extent_writepage(struct bio *bio)
2490 {
2491         struct bio_vec *bvec;
2492         u64 start;
2493         u64 end;
2494         int i;
2495
2496         bio_for_each_segment_all(bvec, bio, i) {
2497                 struct page *page = bvec->bv_page;
2498
2499                 /* We always issue full-page reads, but if some block
2500                  * in a page fails to read, blk_update_request() will
2501                  * advance bv_offset and adjust bv_len to compensate.
2502                  * Print a warning for nonzero offsets, and an error
2503                  * if they don't add up to a full page.  */
2504                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2505                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2506                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2507                                    "partial page write in btrfs with offset %u and length %u",
2508                                         bvec->bv_offset, bvec->bv_len);
2509                         else
2510                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2511                                    "incomplete page write in btrfs with offset %u and "
2512                                    "length %u",
2513                                         bvec->bv_offset, bvec->bv_len);
2514                 }
2515
2516                 start = page_offset(page);
2517                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2518
2519                 if (end_extent_writepage(page, bio->bi_error, start, end))
2520                         continue;
2521
2522                 end_page_writeback(page);
2523         }
2524
2525         bio_put(bio);
2526 }
2527
2528 static void
2529 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2530                               int uptodate)
2531 {
2532         struct extent_state *cached = NULL;
2533         u64 end = start + len - 1;
2534
2535         if (uptodate && tree->track_uptodate)
2536                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2537         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2538 }
2539
2540 /*
2541  * after a readpage IO is done, we need to:
2542  * clear the uptodate bits on error
2543  * set the uptodate bits if things worked
2544  * set the page up to date if all extents in the tree are uptodate
2545  * clear the lock bit in the extent tree
2546  * unlock the page if there are no other extents locked for it
2547  *
2548  * Scheduling is not allowed, so the extent state tree is expected
2549  * to have one and only one object corresponding to this IO.
2550  */
2551 static void end_bio_extent_readpage(struct bio *bio)
2552 {
2553         struct bio_vec *bvec;
2554         int uptodate = !bio->bi_error;
2555         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2556         struct extent_io_tree *tree;
2557         u64 offset = 0;
2558         u64 start;
2559         u64 end;
2560         u64 len;
2561         u64 extent_start = 0;
2562         u64 extent_len = 0;
2563         int mirror;
2564         int ret;
2565         int i;
2566
2567         bio_for_each_segment_all(bvec, bio, i) {
2568                 struct page *page = bvec->bv_page;
2569                 struct inode *inode = page->mapping->host;
2570
2571                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2572                          "mirror=%u\n", (u64)bio->bi_iter.bi_sector,
2573                          bio->bi_error, io_bio->mirror_num);
2574                 tree = &BTRFS_I(inode)->io_tree;
2575
2576                 /* We always issue full-page reads, but if some block
2577                  * in a page fails to read, blk_update_request() will
2578                  * advance bv_offset and adjust bv_len to compensate.
2579                  * Print a warning for nonzero offsets, and an error
2580                  * if they don't add up to a full page.  */
2581                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2582                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2583                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2584                                    "partial page read in btrfs with offset %u and length %u",
2585                                         bvec->bv_offset, bvec->bv_len);
2586                         else
2587                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2588                                    "incomplete page read in btrfs with offset %u and "
2589                                    "length %u",
2590                                         bvec->bv_offset, bvec->bv_len);
2591                 }
2592
2593                 start = page_offset(page);
2594                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2595                 len = bvec->bv_len;
2596
2597                 mirror = io_bio->mirror_num;
2598                 if (likely(uptodate && tree->ops &&
2599                            tree->ops->readpage_end_io_hook)) {
2600                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2601                                                               page, start, end,
2602                                                               mirror);
2603                         if (ret)
2604                                 uptodate = 0;
2605                         else
2606                                 clean_io_failure(inode, start, page, 0);
2607                 }
2608
2609                 if (likely(uptodate))
2610                         goto readpage_ok;
2611
2612                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2613                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2614                         if (!ret && !bio->bi_error)
2615                                 uptodate = 1;
2616                 } else {
2617                         /*
2618                          * The generic bio_readpage_error handles errors the
2619                          * following way: If possible, new read requests are
2620                          * created and submitted and will end up in
2621                          * end_bio_extent_readpage as well (if we're lucky, not
2622                          * in the !uptodate case). In that case it returns 0 and
2623                          * we just go on with the next page in our bio. If it
2624                          * can't handle the error it will return -EIO and we
2625                          * remain responsible for that page.
2626                          */
2627                         ret = bio_readpage_error(bio, offset, page, start, end,
2628                                                  mirror);
2629                         if (ret == 0) {
2630                                 uptodate = !bio->bi_error;
2631                                 offset += len;
2632                                 continue;
2633                         }
2634                 }
2635 readpage_ok:
2636                 if (likely(uptodate)) {
2637                         loff_t i_size = i_size_read(inode);
2638                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2639                         unsigned off;
2640
2641                         /* Zero out the end if this page straddles i_size */
2642                         off = i_size & (PAGE_CACHE_SIZE-1);
2643                         if (page->index == end_index && off)
2644                                 zero_user_segment(page, off, PAGE_CACHE_SIZE);
2645                         SetPageUptodate(page);
2646                 } else {
2647                         ClearPageUptodate(page);
2648                         SetPageError(page);
2649                 }
2650                 unlock_page(page);
2651                 offset += len;
2652
2653                 if (unlikely(!uptodate)) {
2654                         if (extent_len) {
2655                                 endio_readpage_release_extent(tree,
2656                                                               extent_start,
2657                                                               extent_len, 1);
2658                                 extent_start = 0;
2659                                 extent_len = 0;
2660                         }
2661                         endio_readpage_release_extent(tree, start,
2662                                                       end - start + 1, 0);
2663                 } else if (!extent_len) {
2664                         extent_start = start;
2665                         extent_len = end + 1 - start;
2666                 } else if (extent_start + extent_len == start) {
2667                         extent_len += end + 1 - start;
2668                 } else {
2669                         endio_readpage_release_extent(tree, extent_start,
2670                                                       extent_len, uptodate);
2671                         extent_start = start;
2672                         extent_len = end + 1 - start;
2673                 }
2674         }
2675
2676         if (extent_len)
2677                 endio_readpage_release_extent(tree, extent_start, extent_len,
2678                                               uptodate);
2679         if (io_bio->end_io)
2680                 io_bio->end_io(io_bio, bio->bi_error);
2681         bio_put(bio);
2682 }
2683
2684 /*
2685  * this allocates from the btrfs_bioset.  We're returning a bio right now
2686  * but you can call btrfs_io_bio for the appropriate container_of magic
2687  */
2688 struct bio *
2689 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2690                 gfp_t gfp_flags)
2691 {
2692         struct btrfs_io_bio *btrfs_bio;
2693         struct bio *bio;
2694
2695         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2696
2697         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2698                 while (!bio && (nr_vecs /= 2)) {
2699                         bio = bio_alloc_bioset(gfp_flags,
2700                                                nr_vecs, btrfs_bioset);
2701                 }
2702         }
2703
2704         if (bio) {
2705                 bio->bi_bdev = bdev;
2706                 bio->bi_iter.bi_sector = first_sector;
2707                 btrfs_bio = btrfs_io_bio(bio);
2708                 btrfs_bio->csum = NULL;
2709                 btrfs_bio->csum_allocated = NULL;
2710                 btrfs_bio->end_io = NULL;
2711         }
2712         return bio;
2713 }
2714
2715 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2716 {
2717         struct btrfs_io_bio *btrfs_bio;
2718         struct bio *new;
2719
2720         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2721         if (new) {
2722                 btrfs_bio = btrfs_io_bio(new);
2723                 btrfs_bio->csum = NULL;
2724                 btrfs_bio->csum_allocated = NULL;
2725                 btrfs_bio->end_io = NULL;
2726         }
2727         return new;
2728 }
2729
2730 /* this also allocates from the btrfs_bioset */
2731 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2732 {
2733         struct btrfs_io_bio *btrfs_bio;
2734         struct bio *bio;
2735
2736         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2737         if (bio) {
2738                 btrfs_bio = btrfs_io_bio(bio);
2739                 btrfs_bio->csum = NULL;
2740                 btrfs_bio->csum_allocated = NULL;
2741                 btrfs_bio->end_io = NULL;
2742         }
2743         return bio;
2744 }
2745
2746
2747 static int __must_check submit_one_bio(int rw, struct bio *bio,
2748                                        int mirror_num, unsigned long bio_flags)
2749 {
2750         int ret = 0;
2751         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2752         struct page *page = bvec->bv_page;
2753         struct extent_io_tree *tree = bio->bi_private;
2754         u64 start;
2755
2756         start = page_offset(page) + bvec->bv_offset;
2757
2758         bio->bi_private = NULL;
2759
2760         bio_get(bio);
2761
2762         if (tree->ops && tree->ops->submit_bio_hook)
2763                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2764                                            mirror_num, bio_flags, start);
2765         else
2766                 btrfsic_submit_bio(rw, bio);
2767
2768         bio_put(bio);
2769         return ret;
2770 }
2771
2772 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2773                      unsigned long offset, size_t size, struct bio *bio,
2774                      unsigned long bio_flags)
2775 {
2776         int ret = 0;
2777         if (tree->ops && tree->ops->merge_bio_hook)
2778                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2779                                                 bio_flags);
2780         BUG_ON(ret < 0);
2781         return ret;
2782
2783 }
2784
2785 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2786                               struct page *page, sector_t sector,
2787                               size_t size, unsigned long offset,
2788                               struct block_device *bdev,
2789                               struct bio **bio_ret,
2790                               unsigned long max_pages,
2791                               bio_end_io_t end_io_func,
2792                               int mirror_num,
2793                               unsigned long prev_bio_flags,
2794                               unsigned long bio_flags)
2795 {
2796         int ret = 0;
2797         struct bio *bio;
2798         int contig = 0;
2799         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2800         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2801
2802         if (bio_ret && *bio_ret) {
2803                 bio = *bio_ret;
2804                 if (old_compressed)
2805                         contig = bio->bi_iter.bi_sector == sector;
2806                 else
2807                         contig = bio_end_sector(bio) == sector;
2808
2809                 if (prev_bio_flags != bio_flags || !contig ||
2810                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2811                     bio_add_page(bio, page, page_size, offset) < page_size) {
2812                         ret = submit_one_bio(rw, bio, mirror_num,
2813                                              prev_bio_flags);
2814                         if (ret < 0) {
2815                                 *bio_ret = NULL;
2816                                 return ret;
2817                         }
2818                         bio = NULL;
2819                 } else {
2820                         return 0;
2821                 }
2822         }
2823
2824         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2825                         GFP_NOFS | __GFP_HIGH);
2826         if (!bio)
2827                 return -ENOMEM;
2828
2829         bio_add_page(bio, page, page_size, offset);
2830         bio->bi_end_io = end_io_func;
2831         bio->bi_private = tree;
2832
2833         if (bio_ret)
2834                 *bio_ret = bio;
2835         else
2836                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2837
2838         return ret;
2839 }
2840
2841 static void attach_extent_buffer_page(struct extent_buffer *eb,
2842                                       struct page *page)
2843 {
2844         if (!PagePrivate(page)) {
2845                 SetPagePrivate(page);
2846                 page_cache_get(page);
2847                 set_page_private(page, (unsigned long)eb);
2848         } else {
2849                 WARN_ON(page->private != (unsigned long)eb);
2850         }
2851 }
2852
2853 void set_page_extent_mapped(struct page *page)
2854 {
2855         if (!PagePrivate(page)) {
2856                 SetPagePrivate(page);
2857                 page_cache_get(page);
2858                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2859         }
2860 }
2861
2862 static struct extent_map *
2863 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2864                  u64 start, u64 len, get_extent_t *get_extent,
2865                  struct extent_map **em_cached)
2866 {
2867         struct extent_map *em;
2868
2869         if (em_cached && *em_cached) {
2870                 em = *em_cached;
2871                 if (extent_map_in_tree(em) && start >= em->start &&
2872                     start < extent_map_end(em)) {
2873                         atomic_inc(&em->refs);
2874                         return em;
2875                 }
2876
2877                 free_extent_map(em);
2878                 *em_cached = NULL;
2879         }
2880
2881         em = get_extent(inode, page, pg_offset, start, len, 0);
2882         if (em_cached && !IS_ERR_OR_NULL(em)) {
2883                 BUG_ON(*em_cached);
2884                 atomic_inc(&em->refs);
2885                 *em_cached = em;
2886         }
2887         return em;
2888 }
2889 /*
2890  * basic readpage implementation.  Locked extent state structs are inserted
2891  * into the tree that are removed when the IO is done (by the end_io
2892  * handlers)
2893  * XXX JDM: This needs looking at to ensure proper page locking
2894  */
2895 static int __do_readpage(struct extent_io_tree *tree,
2896                          struct page *page,
2897                          get_extent_t *get_extent,
2898                          struct extent_map **em_cached,
2899                          struct bio **bio, int mirror_num,
2900                          unsigned long *bio_flags, int rw)
2901 {
2902         struct inode *inode = page->mapping->host;
2903         u64 start = page_offset(page);
2904         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2905         u64 end;
2906         u64 cur = start;
2907         u64 extent_offset;
2908         u64 last_byte = i_size_read(inode);
2909         u64 block_start;
2910         u64 cur_end;
2911         sector_t sector;
2912         struct extent_map *em;
2913         struct block_device *bdev;
2914         int ret;
2915         int nr = 0;
2916         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2917         size_t pg_offset = 0;
2918         size_t iosize;
2919         size_t disk_io_size;
2920         size_t blocksize = inode->i_sb->s_blocksize;
2921         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2922
2923         set_page_extent_mapped(page);
2924
2925         end = page_end;
2926         if (!PageUptodate(page)) {
2927                 if (cleancache_get_page(page) == 0) {
2928                         BUG_ON(blocksize != PAGE_SIZE);
2929                         unlock_extent(tree, start, end);
2930                         goto out;
2931                 }
2932         }
2933
2934         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2935                 char *userpage;
2936                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2937
2938                 if (zero_offset) {
2939                         iosize = PAGE_CACHE_SIZE - zero_offset;
2940                         userpage = kmap_atomic(page);
2941                         memset(userpage + zero_offset, 0, iosize);
2942                         flush_dcache_page(page);
2943                         kunmap_atomic(userpage);
2944                 }
2945         }
2946         while (cur <= end) {
2947                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2948
2949                 if (cur >= last_byte) {
2950                         char *userpage;
2951                         struct extent_state *cached = NULL;
2952
2953                         iosize = PAGE_CACHE_SIZE - pg_offset;
2954                         userpage = kmap_atomic(page);
2955                         memset(userpage + pg_offset, 0, iosize);
2956                         flush_dcache_page(page);
2957                         kunmap_atomic(userpage);
2958                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2959                                             &cached, GFP_NOFS);
2960                         if (!parent_locked)
2961                                 unlock_extent_cached(tree, cur,
2962                                                      cur + iosize - 1,
2963                                                      &cached, GFP_NOFS);
2964                         break;
2965                 }
2966                 em = __get_extent_map(inode, page, pg_offset, cur,
2967                                       end - cur + 1, get_extent, em_cached);
2968                 if (IS_ERR_OR_NULL(em)) {
2969                         SetPageError(page);
2970                         if (!parent_locked)
2971                                 unlock_extent(tree, cur, end);
2972                         break;
2973                 }
2974                 extent_offset = cur - em->start;
2975                 BUG_ON(extent_map_end(em) <= cur);
2976                 BUG_ON(end < cur);
2977
2978                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2979                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2980                         extent_set_compress_type(&this_bio_flag,
2981                                                  em->compress_type);
2982                 }
2983
2984                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2985                 cur_end = min(extent_map_end(em) - 1, end);
2986                 iosize = ALIGN(iosize, blocksize);
2987                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2988                         disk_io_size = em->block_len;
2989                         sector = em->block_start >> 9;
2990                 } else {
2991                         sector = (em->block_start + extent_offset) >> 9;
2992                         disk_io_size = iosize;
2993                 }
2994                 bdev = em->bdev;
2995                 block_start = em->block_start;
2996                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2997                         block_start = EXTENT_MAP_HOLE;
2998                 free_extent_map(em);
2999                 em = NULL;
3000
3001                 /* we've found a hole, just zero and go on */
3002                 if (block_start == EXTENT_MAP_HOLE) {
3003                         char *userpage;
3004                         struct extent_state *cached = NULL;
3005
3006                         userpage = kmap_atomic(page);
3007                         memset(userpage + pg_offset, 0, iosize);
3008                         flush_dcache_page(page);
3009                         kunmap_atomic(userpage);
3010
3011                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3012                                             &cached, GFP_NOFS);
3013                         unlock_extent_cached(tree, cur, cur + iosize - 1,
3014                                              &cached, GFP_NOFS);
3015                         cur = cur + iosize;
3016                         pg_offset += iosize;
3017                         continue;
3018                 }
3019                 /* the get_extent function already copied into the page */
3020                 if (test_range_bit(tree, cur, cur_end,
3021                                    EXTENT_UPTODATE, 1, NULL)) {
3022                         check_page_uptodate(tree, page);
3023                         if (!parent_locked)
3024                                 unlock_extent(tree, cur, cur + iosize - 1);
3025                         cur = cur + iosize;
3026                         pg_offset += iosize;
3027                         continue;
3028                 }
3029                 /* we have an inline extent but it didn't get marked up
3030                  * to date.  Error out
3031                  */
3032                 if (block_start == EXTENT_MAP_INLINE) {
3033                         SetPageError(page);
3034                         if (!parent_locked)
3035                                 unlock_extent(tree, cur, cur + iosize - 1);
3036                         cur = cur + iosize;
3037                         pg_offset += iosize;
3038                         continue;
3039                 }
3040
3041                 pnr -= page->index;
3042                 ret = submit_extent_page(rw, tree, page,
3043                                          sector, disk_io_size, pg_offset,
3044                                          bdev, bio, pnr,
3045                                          end_bio_extent_readpage, mirror_num,
3046                                          *bio_flags,
3047                                          this_bio_flag);
3048                 if (!ret) {
3049                         nr++;
3050                         *bio_flags = this_bio_flag;
3051                 } else {
3052                         SetPageError(page);
3053                         if (!parent_locked)
3054                                 unlock_extent(tree, cur, cur + iosize - 1);
3055                 }
3056                 cur = cur + iosize;
3057                 pg_offset += iosize;
3058         }
3059 out:
3060         if (!nr) {
3061                 if (!PageError(page))
3062                         SetPageUptodate(page);
3063                 unlock_page(page);
3064         }
3065         return 0;
3066 }
3067
3068 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3069                                              struct page *pages[], int nr_pages,
3070                                              u64 start, u64 end,
3071                                              get_extent_t *get_extent,
3072                                              struct extent_map **em_cached,
3073                                              struct bio **bio, int mirror_num,
3074                                              unsigned long *bio_flags, int rw)
3075 {
3076         struct inode *inode;
3077         struct btrfs_ordered_extent *ordered;
3078         int index;
3079
3080         inode = pages[0]->mapping->host;
3081         while (1) {
3082                 lock_extent(tree, start, end);
3083                 ordered = btrfs_lookup_ordered_range(inode, start,
3084                                                      end - start + 1);
3085                 if (!ordered)
3086                         break;
3087                 unlock_extent(tree, start, end);
3088                 btrfs_start_ordered_extent(inode, ordered, 1);
3089                 btrfs_put_ordered_extent(ordered);
3090         }
3091
3092         for (index = 0; index < nr_pages; index++) {
3093                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3094                               mirror_num, bio_flags, rw);
3095                 page_cache_release(pages[index]);
3096         }
3097 }
3098
3099 static void __extent_readpages(struct extent_io_tree *tree,
3100                                struct page *pages[],
3101                                int nr_pages, get_extent_t *get_extent,
3102                                struct extent_map **em_cached,
3103                                struct bio **bio, int mirror_num,
3104                                unsigned long *bio_flags, int rw)
3105 {
3106         u64 start = 0;
3107         u64 end = 0;
3108         u64 page_start;
3109         int index;
3110         int first_index = 0;
3111
3112         for (index = 0; index < nr_pages; index++) {
3113                 page_start = page_offset(pages[index]);
3114                 if (!end) {
3115                         start = page_start;
3116                         end = start + PAGE_CACHE_SIZE - 1;
3117                         first_index = index;
3118                 } else if (end + 1 == page_start) {
3119                         end += PAGE_CACHE_SIZE;
3120                 } else {
3121                         __do_contiguous_readpages(tree, &pages[first_index],
3122                                                   index - first_index, start,
3123                                                   end, get_extent, em_cached,
3124                                                   bio, mirror_num, bio_flags,
3125                                                   rw);
3126                         start = page_start;
3127                         end = start + PAGE_CACHE_SIZE - 1;
3128                         first_index = index;
3129                 }
3130         }
3131
3132         if (end)
3133                 __do_contiguous_readpages(tree, &pages[first_index],
3134                                           index - first_index, start,
3135                                           end, get_extent, em_cached, bio,
3136                                           mirror_num, bio_flags, rw);
3137 }
3138
3139 static int __extent_read_full_page(struct extent_io_tree *tree,
3140                                    struct page *page,
3141                                    get_extent_t *get_extent,
3142                                    struct bio **bio, int mirror_num,
3143                                    unsigned long *bio_flags, int rw)
3144 {
3145         struct inode *inode = page->mapping->host;
3146         struct btrfs_ordered_extent *ordered;
3147         u64 start = page_offset(page);
3148         u64 end = start + PAGE_CACHE_SIZE - 1;
3149         int ret;
3150
3151         while (1) {
3152                 lock_extent(tree, start, end);
3153                 ordered = btrfs_lookup_ordered_extent(inode, start);
3154                 if (!ordered)
3155                         break;
3156                 unlock_extent(tree, start, end);
3157                 btrfs_start_ordered_extent(inode, ordered, 1);
3158                 btrfs_put_ordered_extent(ordered);
3159         }
3160
3161         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3162                             bio_flags, rw);
3163         return ret;
3164 }
3165
3166 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3167                             get_extent_t *get_extent, int mirror_num)
3168 {
3169         struct bio *bio = NULL;
3170         unsigned long bio_flags = 0;
3171         int ret;
3172
3173         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3174                                       &bio_flags, READ);
3175         if (bio)
3176                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3177         return ret;
3178 }
3179
3180 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3181                                  get_extent_t *get_extent, int mirror_num)
3182 {
3183         struct bio *bio = NULL;
3184         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3185         int ret;
3186
3187         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3188                                       &bio_flags, READ);
3189         if (bio)
3190                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3191         return ret;
3192 }
3193
3194 static noinline void update_nr_written(struct page *page,
3195                                       struct writeback_control *wbc,
3196                                       unsigned long nr_written)
3197 {
3198         wbc->nr_to_write -= nr_written;
3199         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3200             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3201                 page->mapping->writeback_index = page->index + nr_written;
3202 }
3203
3204 /*
3205  * helper for __extent_writepage, doing all of the delayed allocation setup.
3206  *
3207  * This returns 1 if our fill_delalloc function did all the work required
3208  * to write the page (copy into inline extent).  In this case the IO has
3209  * been started and the page is already unlocked.
3210  *
3211  * This returns 0 if all went well (page still locked)
3212  * This returns < 0 if there were errors (page still locked)
3213  */
3214 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3215                               struct page *page, struct writeback_control *wbc,
3216                               struct extent_page_data *epd,
3217                               u64 delalloc_start,
3218                               unsigned long *nr_written)
3219 {
3220         struct extent_io_tree *tree = epd->tree;
3221         u64 page_end = delalloc_start + PAGE_CACHE_SIZE - 1;
3222         u64 nr_delalloc;
3223         u64 delalloc_to_write = 0;
3224         u64 delalloc_end = 0;
3225         int ret;
3226         int page_started = 0;
3227
3228         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3229                 return 0;
3230
3231         while (delalloc_end < page_end) {
3232                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3233                                                page,
3234                                                &delalloc_start,
3235                                                &delalloc_end,
3236                                                BTRFS_MAX_EXTENT_SIZE);
3237                 if (nr_delalloc == 0) {
3238                         delalloc_start = delalloc_end + 1;
3239                         continue;
3240                 }
3241                 ret = tree->ops->fill_delalloc(inode, page,
3242                                                delalloc_start,
3243                                                delalloc_end,
3244                                                &page_started,
3245                                                nr_written);
3246                 /* File system has been set read-only */
3247                 if (ret) {
3248                         SetPageError(page);
3249                         /* fill_delalloc should be return < 0 for error
3250                          * but just in case, we use > 0 here meaning the
3251                          * IO is started, so we don't want to return > 0
3252                          * unless things are going well.
3253                          */
3254                         ret = ret < 0 ? ret : -EIO;
3255                         goto done;
3256                 }
3257                 /*
3258                  * delalloc_end is already one less than the total
3259                  * length, so we don't subtract one from
3260                  * PAGE_CACHE_SIZE
3261                  */
3262                 delalloc_to_write += (delalloc_end - delalloc_start +
3263                                       PAGE_CACHE_SIZE) >>
3264                                       PAGE_CACHE_SHIFT;
3265                 delalloc_start = delalloc_end + 1;
3266         }
3267         if (wbc->nr_to_write < delalloc_to_write) {
3268                 int thresh = 8192;
3269
3270                 if (delalloc_to_write < thresh * 2)
3271                         thresh = delalloc_to_write;
3272                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3273                                          thresh);
3274         }
3275
3276         /* did the fill delalloc function already unlock and start
3277          * the IO?
3278          */
3279         if (page_started) {
3280                 /*
3281                  * we've unlocked the page, so we can't update
3282                  * the mapping's writeback index, just update
3283                  * nr_to_write.
3284                  */
3285                 wbc->nr_to_write -= *nr_written;
3286                 return 1;
3287         }
3288
3289         ret = 0;
3290
3291 done:
3292         return ret;
3293 }
3294
3295 /*
3296  * helper for __extent_writepage.  This calls the writepage start hooks,
3297  * and does the loop to map the page into extents and bios.
3298  *
3299  * We return 1 if the IO is started and the page is unlocked,
3300  * 0 if all went well (page still locked)
3301  * < 0 if there were errors (page still locked)
3302  */
3303 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3304                                  struct page *page,
3305                                  struct writeback_control *wbc,
3306                                  struct extent_page_data *epd,
3307                                  loff_t i_size,
3308                                  unsigned long nr_written,
3309                                  int write_flags, int *nr_ret)
3310 {
3311         struct extent_io_tree *tree = epd->tree;
3312         u64 start = page_offset(page);
3313         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3314         u64 end;
3315         u64 cur = start;
3316         u64 extent_offset;
3317         u64 block_start;
3318         u64 iosize;
3319         sector_t sector;
3320         struct extent_state *cached_state = NULL;
3321         struct extent_map *em;
3322         struct block_device *bdev;
3323         size_t pg_offset = 0;
3324         size_t blocksize;
3325         int ret = 0;
3326         int nr = 0;
3327         bool compressed;
3328
3329         if (tree->ops && tree->ops->writepage_start_hook) {
3330                 ret = tree->ops->writepage_start_hook(page, start,
3331                                                       page_end);
3332                 if (ret) {
3333                         /* Fixup worker will requeue */
3334                         if (ret == -EBUSY)
3335                                 wbc->pages_skipped++;
3336                         else
3337                                 redirty_page_for_writepage(wbc, page);
3338
3339                         update_nr_written(page, wbc, nr_written);
3340                         unlock_page(page);
3341                         ret = 1;
3342                         goto done_unlocked;
3343                 }
3344         }
3345
3346         /*
3347          * we don't want to touch the inode after unlocking the page,
3348          * so we update the mapping writeback index now
3349          */
3350         update_nr_written(page, wbc, nr_written + 1);
3351
3352         end = page_end;
3353         if (i_size <= start) {
3354                 if (tree->ops && tree->ops->writepage_end_io_hook)
3355                         tree->ops->writepage_end_io_hook(page, start,
3356                                                          page_end, NULL, 1);
3357                 goto done;
3358         }
3359
3360         blocksize = inode->i_sb->s_blocksize;
3361
3362         while (cur <= end) {
3363                 u64 em_end;
3364                 if (cur >= i_size) {
3365                         if (tree->ops && tree->ops->writepage_end_io_hook)
3366                                 tree->ops->writepage_end_io_hook(page, cur,
3367                                                          page_end, NULL, 1);
3368                         break;
3369                 }
3370                 em = epd->get_extent(inode, page, pg_offset, cur,
3371                                      end - cur + 1, 1);
3372                 if (IS_ERR_OR_NULL(em)) {
3373                         SetPageError(page);
3374                         ret = PTR_ERR_OR_ZERO(em);
3375                         break;
3376                 }
3377
3378                 extent_offset = cur - em->start;
3379                 em_end = extent_map_end(em);
3380                 BUG_ON(em_end <= cur);
3381                 BUG_ON(end < cur);
3382                 iosize = min(em_end - cur, end - cur + 1);
3383                 iosize = ALIGN(iosize, blocksize);
3384                 sector = (em->block_start + extent_offset) >> 9;
3385                 bdev = em->bdev;
3386                 block_start = em->block_start;
3387                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3388                 free_extent_map(em);
3389                 em = NULL;
3390
3391                 /*
3392                  * compressed and inline extents are written through other
3393                  * paths in the FS
3394                  */
3395                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3396                     block_start == EXTENT_MAP_INLINE) {
3397                         /*
3398                          * end_io notification does not happen here for
3399                          * compressed extents
3400                          */
3401                         if (!compressed && tree->ops &&
3402                             tree->ops->writepage_end_io_hook)
3403                                 tree->ops->writepage_end_io_hook(page, cur,
3404                                                          cur + iosize - 1,
3405                                                          NULL, 1);
3406                         else if (compressed) {
3407                                 /* we don't want to end_page_writeback on
3408                                  * a compressed extent.  this happens
3409                                  * elsewhere
3410                                  */
3411                                 nr++;
3412                         }
3413
3414                         cur += iosize;
3415                         pg_offset += iosize;
3416                         continue;
3417                 }
3418
3419                 if (tree->ops && tree->ops->writepage_io_hook) {
3420                         ret = tree->ops->writepage_io_hook(page, cur,
3421                                                 cur + iosize - 1);
3422                 } else {
3423                         ret = 0;
3424                 }
3425                 if (ret) {
3426                         SetPageError(page);
3427                 } else {
3428                         unsigned long max_nr = (i_size >> PAGE_CACHE_SHIFT) + 1;
3429
3430                         set_range_writeback(tree, cur, cur + iosize - 1);
3431                         if (!PageWriteback(page)) {
3432                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
3433                                            "page %lu not writeback, cur %llu end %llu",
3434                                        page->index, cur, end);
3435                         }
3436
3437                         ret = submit_extent_page(write_flags, tree, page,
3438                                                  sector, iosize, pg_offset,
3439                                                  bdev, &epd->bio, max_nr,
3440                                                  end_bio_extent_writepage,
3441                                                  0, 0, 0);
3442                         if (ret)
3443                                 SetPageError(page);
3444                 }
3445                 cur = cur + iosize;
3446                 pg_offset += iosize;
3447                 nr++;
3448         }
3449 done:
3450         *nr_ret = nr;
3451
3452 done_unlocked:
3453
3454         /* drop our reference on any cached states */
3455         free_extent_state(cached_state);
3456         return ret;
3457 }
3458
3459 /*
3460  * the writepage semantics are similar to regular writepage.  extent
3461  * records are inserted to lock ranges in the tree, and as dirty areas
3462  * are found, they are marked writeback.  Then the lock bits are removed
3463  * and the end_io handler clears the writeback ranges
3464  */
3465 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3466                               void *data)
3467 {
3468         struct inode *inode = page->mapping->host;
3469         struct extent_page_data *epd = data;
3470         u64 start = page_offset(page);
3471         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3472         int ret;
3473         int nr = 0;
3474         size_t pg_offset = 0;
3475         loff_t i_size = i_size_read(inode);
3476         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3477         int write_flags;
3478         unsigned long nr_written = 0;
3479
3480         if (wbc->sync_mode == WB_SYNC_ALL)
3481                 write_flags = WRITE_SYNC;
3482         else
3483                 write_flags = WRITE;
3484
3485         trace___extent_writepage(page, inode, wbc);
3486
3487         WARN_ON(!PageLocked(page));
3488
3489         ClearPageError(page);
3490
3491         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3492         if (page->index > end_index ||
3493            (page->index == end_index && !pg_offset)) {
3494                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3495                 unlock_page(page);
3496                 return 0;
3497         }
3498
3499         if (page->index == end_index) {
3500                 char *userpage;
3501
3502                 userpage = kmap_atomic(page);
3503                 memset(userpage + pg_offset, 0,
3504                        PAGE_CACHE_SIZE - pg_offset);
3505                 kunmap_atomic(userpage);
3506                 flush_dcache_page(page);
3507         }
3508
3509         pg_offset = 0;
3510
3511         set_page_extent_mapped(page);
3512
3513         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3514         if (ret == 1)
3515                 goto done_unlocked;
3516         if (ret)
3517                 goto done;
3518
3519         ret = __extent_writepage_io(inode, page, wbc, epd,
3520                                     i_size, nr_written, write_flags, &nr);
3521         if (ret == 1)
3522                 goto done_unlocked;
3523
3524 done:
3525         if (nr == 0) {
3526                 /* make sure the mapping tag for page dirty gets cleared */
3527                 set_page_writeback(page);
3528                 end_page_writeback(page);
3529         }
3530         if (PageError(page)) {
3531                 ret = ret < 0 ? ret : -EIO;
3532                 end_extent_writepage(page, ret, start, page_end);
3533         }
3534         unlock_page(page);
3535         return ret;
3536
3537 done_unlocked:
3538         return 0;
3539 }
3540
3541 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3542 {
3543         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3544                        TASK_UNINTERRUPTIBLE);
3545 }
3546
3547 static noinline_for_stack int
3548 lock_extent_buffer_for_io(struct extent_buffer *eb,
3549                           struct btrfs_fs_info *fs_info,
3550                           struct extent_page_data *epd)
3551 {
3552         unsigned long i, num_pages;
3553         int flush = 0;
3554         int ret = 0;
3555
3556         if (!btrfs_try_tree_write_lock(eb)) {
3557                 flush = 1;
3558                 flush_write_bio(epd);
3559                 btrfs_tree_lock(eb);
3560         }
3561
3562         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3563                 btrfs_tree_unlock(eb);
3564                 if (!epd->sync_io)
3565                         return 0;
3566                 if (!flush) {
3567                         flush_write_bio(epd);
3568                         flush = 1;
3569                 }
3570                 while (1) {
3571                         wait_on_extent_buffer_writeback(eb);
3572                         btrfs_tree_lock(eb);
3573                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3574                                 break;
3575                         btrfs_tree_unlock(eb);
3576                 }
3577         }
3578
3579         /*
3580          * We need to do this to prevent races in people who check if the eb is
3581          * under IO since we can end up having no IO bits set for a short period
3582          * of time.
3583          */
3584         spin_lock(&eb->refs_lock);
3585         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3586                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3587                 spin_unlock(&eb->refs_lock);
3588                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3589                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3590                                      -eb->len,
3591                                      fs_info->dirty_metadata_batch);
3592                 ret = 1;
3593         } else {
3594                 spin_unlock(&eb->refs_lock);
3595         }
3596
3597         btrfs_tree_unlock(eb);
3598
3599         if (!ret)
3600                 return ret;
3601
3602         num_pages = num_extent_pages(eb->start, eb->len);
3603         for (i = 0; i < num_pages; i++) {
3604                 struct page *p = eb->pages[i];
3605
3606                 if (!trylock_page(p)) {
3607                         if (!flush) {
3608                                 flush_write_bio(epd);
3609                                 flush = 1;
3610                         }
3611                         lock_page(p);
3612                 }
3613         }
3614
3615         return ret;
3616 }
3617
3618 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3619 {
3620         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3621         smp_mb__after_atomic();
3622         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3623 }
3624
3625 static void set_btree_ioerr(struct page *page)
3626 {
3627         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3628         struct btrfs_inode *btree_ino = BTRFS_I(eb->fs_info->btree_inode);
3629
3630         SetPageError(page);
3631         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3632                 return;
3633
3634         /*
3635          * If writeback for a btree extent that doesn't belong to a log tree
3636          * failed, increment the counter transaction->eb_write_errors.
3637          * We do this because while the transaction is running and before it's
3638          * committing (when we call filemap_fdata[write|wait]_range against
3639          * the btree inode), we might have
3640          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3641          * returns an error or an error happens during writeback, when we're
3642          * committing the transaction we wouldn't know about it, since the pages
3643          * can be no longer dirty nor marked anymore for writeback (if a
3644          * subsequent modification to the extent buffer didn't happen before the
3645          * transaction commit), which makes filemap_fdata[write|wait]_range not
3646          * able to find the pages tagged with SetPageError at transaction
3647          * commit time. So if this happens we must abort the transaction,
3648          * otherwise we commit a super block with btree roots that point to
3649          * btree nodes/leafs whose content on disk is invalid - either garbage
3650          * or the content of some node/leaf from a past generation that got
3651          * cowed or deleted and is no longer valid.
3652          *
3653          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3654          * not be enough - we need to distinguish between log tree extents vs
3655          * non-log tree extents, and the next filemap_fdatawait_range() call
3656          * will catch and clear such errors in the mapping - and that call might
3657          * be from a log sync and not from a transaction commit. Also, checking
3658          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3659          * not done and would not be reliable - the eb might have been released
3660          * from memory and reading it back again means that flag would not be
3661          * set (since it's a runtime flag, not persisted on disk).
3662          *
3663          * Using the flags below in the btree inode also makes us achieve the
3664          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3665          * writeback for all dirty pages and before filemap_fdatawait_range()
3666          * is called, the writeback for all dirty pages had already finished
3667          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3668          * filemap_fdatawait_range() would return success, as it could not know
3669          * that writeback errors happened (the pages were no longer tagged for
3670          * writeback).
3671          */
3672         switch (eb->log_index) {
3673         case -1:
3674                 set_bit(BTRFS_INODE_BTREE_ERR, &btree_ino->runtime_flags);
3675                 break;
3676         case 0:
3677                 set_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
3678                 break;
3679         case 1:
3680                 set_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
3681                 break;
3682         default:
3683                 BUG(); /* unexpected, logic error */
3684         }
3685 }
3686
3687 static void end_bio_extent_buffer_writepage(struct bio *bio)
3688 {
3689         struct bio_vec *bvec;
3690         struct extent_buffer *eb;
3691         int i, done;
3692
3693         bio_for_each_segment_all(bvec, bio, i) {
3694                 struct page *page = bvec->bv_page;
3695
3696                 eb = (struct extent_buffer *)page->private;
3697                 BUG_ON(!eb);
3698                 done = atomic_dec_and_test(&eb->io_pages);
3699
3700                 if (bio->bi_error ||
3701                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3702                         ClearPageUptodate(page);
3703                         set_btree_ioerr(page);
3704                 }
3705
3706                 end_page_writeback(page);
3707
3708                 if (!done)
3709                         continue;
3710
3711                 end_extent_buffer_writeback(eb);
3712         }
3713
3714         bio_put(bio);
3715 }
3716
3717 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3718                         struct btrfs_fs_info *fs_info,
3719                         struct writeback_control *wbc,
3720                         struct extent_page_data *epd)
3721 {
3722         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3723         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3724         u64 offset = eb->start;
3725         unsigned long i, num_pages;
3726         unsigned long bio_flags = 0;
3727         int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3728         int ret = 0;
3729
3730         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3731         num_pages = num_extent_pages(eb->start, eb->len);
3732         atomic_set(&eb->io_pages, num_pages);
3733         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3734                 bio_flags = EXTENT_BIO_TREE_LOG;
3735
3736         for (i = 0; i < num_pages; i++) {
3737                 struct page *p = eb->pages[i];
3738
3739                 clear_page_dirty_for_io(p);
3740                 set_page_writeback(p);
3741                 ret = submit_extent_page(rw, tree, p, offset >> 9,
3742                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3743                                          -1, end_bio_extent_buffer_writepage,
3744                                          0, epd->bio_flags, bio_flags);
3745                 epd->bio_flags = bio_flags;
3746                 if (ret) {
3747                         set_btree_ioerr(p);
3748                         end_page_writeback(p);
3749                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3750                                 end_extent_buffer_writeback(eb);
3751                         ret = -EIO;
3752                         break;
3753                 }
3754                 offset += PAGE_CACHE_SIZE;
3755                 update_nr_written(p, wbc, 1);
3756                 unlock_page(p);
3757         }
3758
3759         if (unlikely(ret)) {
3760                 for (; i < num_pages; i++) {
3761                         struct page *p = eb->pages[i];
3762                         clear_page_dirty_for_io(p);
3763                         unlock_page(p);
3764                 }
3765         }
3766
3767         return ret;
3768 }
3769
3770 int btree_write_cache_pages(struct address_space *mapping,
3771                                    struct writeback_control *wbc)
3772 {
3773         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3774         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3775         struct extent_buffer *eb, *prev_eb = NULL;
3776         struct extent_page_data epd = {
3777                 .bio = NULL,
3778                 .tree = tree,
3779                 .extent_locked = 0,
3780                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3781                 .bio_flags = 0,
3782         };
3783         int ret = 0;
3784         int done = 0;
3785         int nr_to_write_done = 0;
3786         struct pagevec pvec;
3787         int nr_pages;
3788         pgoff_t index;
3789         pgoff_t end;            /* Inclusive */
3790         int scanned = 0;
3791         int tag;
3792
3793         pagevec_init(&pvec, 0);
3794         if (wbc->range_cyclic) {
3795                 index = mapping->writeback_index; /* Start from prev offset */
3796                 end = -1;
3797         } else {
3798                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3799                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3800                 scanned = 1;
3801         }
3802         if (wbc->sync_mode == WB_SYNC_ALL)
3803                 tag = PAGECACHE_TAG_TOWRITE;
3804         else
3805                 tag = PAGECACHE_TAG_DIRTY;
3806 retry:
3807         if (wbc->sync_mode == WB_SYNC_ALL)
3808                 tag_pages_for_writeback(mapping, index, end);
3809         while (!done && !nr_to_write_done && (index <= end) &&
3810                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3811                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3812                 unsigned i;
3813
3814                 scanned = 1;
3815                 for (i = 0; i < nr_pages; i++) {
3816                         struct page *page = pvec.pages[i];
3817
3818                         if (!PagePrivate(page))
3819                                 continue;
3820
3821                         if (!wbc->range_cyclic && page->index > end) {
3822                                 done = 1;
3823                                 break;
3824                         }
3825
3826                         spin_lock(&mapping->private_lock);
3827                         if (!PagePrivate(page)) {
3828                                 spin_unlock(&mapping->private_lock);
3829                                 continue;
3830                         }
3831
3832                         eb = (struct extent_buffer *)page->private;
3833
3834                         /*
3835                          * Shouldn't happen and normally this would be a BUG_ON
3836                          * but no sense in crashing the users box for something
3837                          * we can survive anyway.
3838                          */
3839                         if (WARN_ON(!eb)) {
3840                                 spin_unlock(&mapping->private_lock);
3841                                 continue;
3842                         }
3843
3844                         if (eb == prev_eb) {
3845                                 spin_unlock(&mapping->private_lock);
3846                                 continue;
3847                         }
3848
3849                         ret = atomic_inc_not_zero(&eb->refs);
3850                         spin_unlock(&mapping->private_lock);
3851                         if (!ret)
3852                                 continue;
3853
3854                         prev_eb = eb;
3855                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3856                         if (!ret) {
3857                                 free_extent_buffer(eb);
3858                                 continue;
3859                         }
3860
3861                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3862                         if (ret) {
3863                                 done = 1;
3864                                 free_extent_buffer(eb);
3865                                 break;
3866                         }
3867                         free_extent_buffer(eb);
3868
3869                         /*
3870                          * the filesystem may choose to bump up nr_to_write.
3871                          * We have to make sure to honor the new nr_to_write
3872                          * at any time
3873                          */
3874                         nr_to_write_done = wbc->nr_to_write <= 0;
3875                 }
3876                 pagevec_release(&pvec);
3877                 cond_resched();
3878         }
3879         if (!scanned && !done) {
3880                 /*
3881                  * We hit the last page and there is more work to be done: wrap
3882                  * back to the start of the file
3883                  */
3884                 scanned = 1;
3885                 index = 0;
3886                 goto retry;
3887         }
3888         flush_write_bio(&epd);
3889         return ret;
3890 }
3891
3892 /**
3893  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3894  * @mapping: address space structure to write
3895  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3896  * @writepage: function called for each page
3897  * @data: data passed to writepage function
3898  *
3899  * If a page is already under I/O, write_cache_pages() skips it, even
3900  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3901  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3902  * and msync() need to guarantee that all the data which was dirty at the time
3903  * the call was made get new I/O started against them.  If wbc->sync_mode is
3904  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3905  * existing IO to complete.
3906  */
3907 static int extent_write_cache_pages(struct extent_io_tree *tree,
3908                              struct address_space *mapping,
3909                              struct writeback_control *wbc,
3910                              writepage_t writepage, void *data,
3911                              void (*flush_fn)(void *))
3912 {
3913         struct inode *inode = mapping->host;
3914         int ret = 0;
3915         int done = 0;
3916         int err = 0;
3917         int nr_to_write_done = 0;
3918         struct pagevec pvec;
3919         int nr_pages;
3920         pgoff_t index;
3921         pgoff_t end;            /* Inclusive */
3922         int scanned = 0;
3923         int tag;
3924
3925         /*
3926          * We have to hold onto the inode so that ordered extents can do their
3927          * work when the IO finishes.  The alternative to this is failing to add
3928          * an ordered extent if the igrab() fails there and that is a huge pain
3929          * to deal with, so instead just hold onto the inode throughout the
3930          * writepages operation.  If it fails here we are freeing up the inode
3931          * anyway and we'd rather not waste our time writing out stuff that is
3932          * going to be truncated anyway.
3933          */
3934         if (!igrab(inode))
3935                 return 0;
3936
3937         pagevec_init(&pvec, 0);
3938         if (wbc->range_cyclic) {
3939                 index = mapping->writeback_index; /* Start from prev offset */
3940                 end = -1;
3941         } else {
3942                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3943                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3944                 scanned = 1;
3945         }
3946         if (wbc->sync_mode == WB_SYNC_ALL)
3947                 tag = PAGECACHE_TAG_TOWRITE;
3948         else
3949                 tag = PAGECACHE_TAG_DIRTY;
3950 retry:
3951         if (wbc->sync_mode == WB_SYNC_ALL)
3952                 tag_pages_for_writeback(mapping, index, end);
3953         while (!done && !nr_to_write_done && (index <= end) &&
3954                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3955                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3956                 unsigned i;
3957
3958                 scanned = 1;
3959                 for (i = 0; i < nr_pages; i++) {
3960                         struct page *page = pvec.pages[i];
3961
3962                         /*
3963                          * At this point we hold neither mapping->tree_lock nor
3964                          * lock on the page itself: the page may be truncated or
3965                          * invalidated (changing page->mapping to NULL), or even
3966                          * swizzled back from swapper_space to tmpfs file
3967                          * mapping
3968                          */
3969                         if (!trylock_page(page)) {
3970                                 flush_fn(data);
3971                                 lock_page(page);
3972                         }
3973
3974                         if (unlikely(page->mapping != mapping)) {
3975                                 unlock_page(page);
3976                                 continue;
3977                         }
3978
3979                         if (!wbc->range_cyclic && page->index > end) {
3980                                 done = 1;
3981                                 unlock_page(page);
3982                                 continue;
3983                         }
3984
3985                         if (wbc->sync_mode != WB_SYNC_NONE) {
3986                                 if (PageWriteback(page))
3987                                         flush_fn(data);
3988                                 wait_on_page_writeback(page);
3989                         }
3990
3991                         if (PageWriteback(page) ||
3992                             !clear_page_dirty_for_io(page)) {
3993                                 unlock_page(page);
3994                                 continue;
3995                         }
3996
3997                         ret = (*writepage)(page, wbc, data);
3998
3999                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4000                                 unlock_page(page);
4001                                 ret = 0;
4002                         }
4003                         if (!err && ret < 0)
4004                                 err = ret;
4005
4006                         /*
4007                          * the filesystem may choose to bump up nr_to_write.
4008                          * We have to make sure to honor the new nr_to_write
4009                          * at any time
4010                          */
4011                         nr_to_write_done = wbc->nr_to_write <= 0;
4012                 }
4013                 pagevec_release(&pvec);
4014                 cond_resched();
4015         }
4016         if (!scanned && !done && !err) {
4017                 /*
4018                  * We hit the last page and there is more work to be done: wrap
4019                  * back to the start of the file
4020                  */
4021                 scanned = 1;
4022                 index = 0;
4023                 goto retry;
4024         }
4025         btrfs_add_delayed_iput(inode);
4026         return err;
4027 }
4028
4029 static void flush_epd_write_bio(struct extent_page_data *epd)
4030 {
4031         if (epd->bio) {
4032                 int rw = WRITE;
4033                 int ret;
4034
4035                 if (epd->sync_io)
4036                         rw = WRITE_SYNC;
4037
4038                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
4039                 BUG_ON(ret < 0); /* -ENOMEM */
4040                 epd->bio = NULL;
4041         }
4042 }
4043
4044 static noinline void flush_write_bio(void *data)
4045 {
4046         struct extent_page_data *epd = data;
4047         flush_epd_write_bio(epd);
4048 }
4049
4050 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4051                           get_extent_t *get_extent,
4052                           struct writeback_control *wbc)
4053 {
4054         int ret;
4055         struct extent_page_data epd = {
4056                 .bio = NULL,
4057                 .tree = tree,
4058                 .get_extent = get_extent,
4059                 .extent_locked = 0,
4060                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4061                 .bio_flags = 0,
4062         };
4063
4064         ret = __extent_writepage(page, wbc, &epd);
4065
4066         flush_epd_write_bio(&epd);
4067         return ret;
4068 }
4069
4070 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4071                               u64 start, u64 end, get_extent_t *get_extent,
4072                               int mode)
4073 {
4074         int ret = 0;
4075         struct address_space *mapping = inode->i_mapping;
4076         struct page *page;
4077         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
4078                 PAGE_CACHE_SHIFT;
4079
4080         struct extent_page_data epd = {
4081                 .bio = NULL,
4082                 .tree = tree,
4083                 .get_extent = get_extent,
4084                 .extent_locked = 1,
4085                 .sync_io = mode == WB_SYNC_ALL,
4086                 .bio_flags = 0,
4087         };
4088         struct writeback_control wbc_writepages = {
4089                 .sync_mode      = mode,
4090                 .nr_to_write    = nr_pages * 2,
4091                 .range_start    = start,
4092                 .range_end      = end + 1,
4093         };
4094
4095         while (start <= end) {
4096                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
4097                 if (clear_page_dirty_for_io(page))
4098                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4099                 else {
4100                         if (tree->ops && tree->ops->writepage_end_io_hook)
4101                                 tree->ops->writepage_end_io_hook(page, start,
4102                                                  start + PAGE_CACHE_SIZE - 1,
4103                                                  NULL, 1);
4104                         unlock_page(page);
4105                 }
4106                 page_cache_release(page);
4107                 start += PAGE_CACHE_SIZE;
4108         }
4109
4110         flush_epd_write_bio(&epd);
4111         return ret;
4112 }
4113
4114 int extent_writepages(struct extent_io_tree *tree,
4115                       struct address_space *mapping,
4116                       get_extent_t *get_extent,
4117                       struct writeback_control *wbc)
4118 {
4119         int ret = 0;
4120         struct extent_page_data epd = {
4121                 .bio = NULL,
4122                 .tree = tree,
4123                 .get_extent = get_extent,
4124                 .extent_locked = 0,
4125                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4126                 .bio_flags = 0,
4127         };
4128
4129         ret = extent_write_cache_pages(tree, mapping, wbc,
4130                                        __extent_writepage, &epd,
4131                                        flush_write_bio);
4132         flush_epd_write_bio(&epd);
4133         return ret;
4134 }
4135
4136 int extent_readpages(struct extent_io_tree *tree,
4137                      struct address_space *mapping,
4138                      struct list_head *pages, unsigned nr_pages,
4139                      get_extent_t get_extent)
4140 {
4141         struct bio *bio = NULL;
4142         unsigned page_idx;
4143         unsigned long bio_flags = 0;
4144         struct page *pagepool[16];
4145         struct page *page;
4146         struct extent_map *em_cached = NULL;
4147         int nr = 0;
4148
4149         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4150                 page = list_entry(pages->prev, struct page, lru);
4151
4152                 prefetchw(&page->flags);
4153                 list_del(&page->lru);
4154                 if (add_to_page_cache_lru(page, mapping,
4155                                         page->index, GFP_NOFS)) {
4156                         page_cache_release(page);
4157                         continue;
4158                 }
4159
4160                 pagepool[nr++] = page;
4161                 if (nr < ARRAY_SIZE(pagepool))
4162                         continue;
4163                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4164                                    &bio, 0, &bio_flags, READ);
4165                 nr = 0;
4166         }
4167         if (nr)
4168                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4169                                    &bio, 0, &bio_flags, READ);
4170
4171         if (em_cached)
4172                 free_extent_map(em_cached);
4173
4174         BUG_ON(!list_empty(pages));
4175         if (bio)
4176                 return submit_one_bio(READ, bio, 0, bio_flags);
4177         return 0;
4178 }
4179
4180 /*
4181  * basic invalidatepage code, this waits on any locked or writeback
4182  * ranges corresponding to the page, and then deletes any extent state
4183  * records from the tree
4184  */
4185 int extent_invalidatepage(struct extent_io_tree *tree,
4186                           struct page *page, unsigned long offset)
4187 {
4188         struct extent_state *cached_state = NULL;
4189         u64 start = page_offset(page);
4190         u64 end = start + PAGE_CACHE_SIZE - 1;
4191         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4192
4193         start += ALIGN(offset, blocksize);
4194         if (start > end)
4195                 return 0;
4196
4197         lock_extent_bits(tree, start, end, 0, &cached_state);
4198         wait_on_page_writeback(page);
4199         clear_extent_bit(tree, start, end,
4200                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4201                          EXTENT_DO_ACCOUNTING,
4202                          1, 1, &cached_state, GFP_NOFS);
4203         return 0;
4204 }
4205
4206 /*
4207  * a helper for releasepage, this tests for areas of the page that
4208  * are locked or under IO and drops the related state bits if it is safe
4209  * to drop the page.
4210  */
4211 static int try_release_extent_state(struct extent_map_tree *map,
4212                                     struct extent_io_tree *tree,
4213                                     struct page *page, gfp_t mask)
4214 {
4215         u64 start = page_offset(page);
4216         u64 end = start + PAGE_CACHE_SIZE - 1;
4217         int ret = 1;
4218
4219         if (test_range_bit(tree, start, end,
4220                            EXTENT_IOBITS, 0, NULL))
4221                 ret = 0;
4222         else {
4223                 if ((mask & GFP_NOFS) == GFP_NOFS)
4224                         mask = GFP_NOFS;
4225                 /*
4226                  * at this point we can safely clear everything except the
4227                  * locked bit and the nodatasum bit
4228                  */
4229                 ret = clear_extent_bit(tree, start, end,
4230                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4231                                  0, 0, NULL, mask);
4232
4233                 /* if clear_extent_bit failed for enomem reasons,
4234                  * we can't allow the release to continue.
4235                  */
4236                 if (ret < 0)
4237                         ret = 0;
4238                 else
4239                         ret = 1;
4240         }
4241         return ret;
4242 }
4243
4244 /*
4245  * a helper for releasepage.  As long as there are no locked extents
4246  * in the range corresponding to the page, both state records and extent
4247  * map records are removed
4248  */
4249 int try_release_extent_mapping(struct extent_map_tree *map,
4250                                struct extent_io_tree *tree, struct page *page,
4251                                gfp_t mask)
4252 {
4253         struct extent_map *em;
4254         u64 start = page_offset(page);
4255         u64 end = start + PAGE_CACHE_SIZE - 1;
4256
4257         if ((mask & __GFP_WAIT) &&
4258             page->mapping->host->i_size > 16 * 1024 * 1024) {
4259                 u64 len;
4260                 while (start <= end) {
4261                         len = end - start + 1;
4262                         write_lock(&map->lock);
4263                         em = lookup_extent_mapping(map, start, len);
4264                         if (!em) {
4265                                 write_unlock(&map->lock);
4266                                 break;
4267                         }
4268                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4269                             em->start != start) {
4270                                 write_unlock(&map->lock);
4271                                 free_extent_map(em);
4272                                 break;
4273                         }
4274                         if (!test_range_bit(tree, em->start,
4275                                             extent_map_end(em) - 1,
4276                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4277                                             0, NULL)) {
4278                                 remove_extent_mapping(map, em);
4279                                 /* once for the rb tree */
4280                                 free_extent_map(em);
4281                         }
4282                         start = extent_map_end(em);
4283                         write_unlock(&map->lock);
4284
4285                         /* once for us */
4286                         free_extent_map(em);
4287                 }
4288         }
4289         return try_release_extent_state(map, tree, page, mask);
4290 }
4291
4292 /*
4293  * helper function for fiemap, which doesn't want to see any holes.
4294  * This maps until we find something past 'last'
4295  */
4296 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4297                                                 u64 offset,
4298                                                 u64 last,
4299                                                 get_extent_t *get_extent)
4300 {
4301         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4302         struct extent_map *em;
4303         u64 len;
4304
4305         if (offset >= last)
4306                 return NULL;
4307
4308         while (1) {
4309                 len = last - offset;
4310                 if (len == 0)
4311                         break;
4312                 len = ALIGN(len, sectorsize);
4313                 em = get_extent(inode, NULL, 0, offset, len, 0);
4314                 if (IS_ERR_OR_NULL(em))
4315                         return em;
4316
4317                 /* if this isn't a hole return it */
4318                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4319                     em->block_start != EXTENT_MAP_HOLE) {
4320                         return em;
4321                 }
4322
4323                 /* this is a hole, advance to the next extent */
4324                 offset = extent_map_end(em);
4325                 free_extent_map(em);
4326                 if (offset >= last)
4327                         break;
4328         }
4329         return NULL;
4330 }
4331
4332 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4333                 __u64 start, __u64 len, get_extent_t *get_extent)
4334 {
4335         int ret = 0;
4336         u64 off = start;
4337         u64 max = start + len;
4338         u32 flags = 0;
4339         u32 found_type;
4340         u64 last;
4341         u64 last_for_get_extent = 0;
4342         u64 disko = 0;
4343         u64 isize = i_size_read(inode);
4344         struct btrfs_key found_key;
4345         struct extent_map *em = NULL;
4346         struct extent_state *cached_state = NULL;
4347         struct btrfs_path *path;
4348         struct btrfs_root *root = BTRFS_I(inode)->root;
4349         int end = 0;
4350         u64 em_start = 0;
4351         u64 em_len = 0;
4352         u64 em_end = 0;
4353
4354         if (len == 0)
4355                 return -EINVAL;
4356
4357         path = btrfs_alloc_path();
4358         if (!path)
4359                 return -ENOMEM;
4360         path->leave_spinning = 1;
4361
4362         start = round_down(start, BTRFS_I(inode)->root->sectorsize);
4363         len = round_up(max, BTRFS_I(inode)->root->sectorsize) - start;
4364
4365         /*
4366          * lookup the last file extent.  We're not using i_size here
4367          * because there might be preallocation past i_size
4368          */
4369         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode), -1,
4370                                        0);
4371         if (ret < 0) {
4372                 btrfs_free_path(path);
4373                 return ret;
4374         }
4375         WARN_ON(!ret);
4376         path->slots[0]--;
4377         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4378         found_type = found_key.type;
4379
4380         /* No extents, but there might be delalloc bits */
4381         if (found_key.objectid != btrfs_ino(inode) ||
4382             found_type != BTRFS_EXTENT_DATA_KEY) {
4383                 /* have to trust i_size as the end */
4384                 last = (u64)-1;
4385                 last_for_get_extent = isize;
4386         } else {
4387                 /*
4388                  * remember the start of the last extent.  There are a
4389                  * bunch of different factors that go into the length of the
4390                  * extent, so its much less complex to remember where it started
4391                  */
4392                 last = found_key.offset;
4393                 last_for_get_extent = last + 1;
4394         }
4395         btrfs_release_path(path);
4396
4397         /*
4398          * we might have some extents allocated but more delalloc past those
4399          * extents.  so, we trust isize unless the start of the last extent is
4400          * beyond isize
4401          */
4402         if (last < isize) {
4403                 last = (u64)-1;
4404                 last_for_get_extent = isize;
4405         }
4406
4407         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4408                          &cached_state);
4409
4410         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4411                                    get_extent);
4412         if (!em)
4413                 goto out;
4414         if (IS_ERR(em)) {
4415                 ret = PTR_ERR(em);
4416                 goto out;
4417         }
4418
4419         while (!end) {
4420                 u64 offset_in_extent = 0;
4421
4422                 /* break if the extent we found is outside the range */
4423                 if (em->start >= max || extent_map_end(em) < off)
4424                         break;
4425
4426                 /*
4427                  * get_extent may return an extent that starts before our
4428                  * requested range.  We have to make sure the ranges
4429                  * we return to fiemap always move forward and don't
4430                  * overlap, so adjust the offsets here
4431                  */
4432                 em_start = max(em->start, off);
4433
4434                 /*
4435                  * record the offset from the start of the extent
4436                  * for adjusting the disk offset below.  Only do this if the
4437                  * extent isn't compressed since our in ram offset may be past
4438                  * what we have actually allocated on disk.
4439                  */
4440                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4441                         offset_in_extent = em_start - em->start;
4442                 em_end = extent_map_end(em);
4443                 em_len = em_end - em_start;
4444                 disko = 0;
4445                 flags = 0;
4446
4447                 /*
4448                  * bump off for our next call to get_extent
4449                  */
4450                 off = extent_map_end(em);
4451                 if (off >= max)
4452                         end = 1;
4453
4454                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4455                         end = 1;
4456                         flags |= FIEMAP_EXTENT_LAST;
4457                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4458                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4459                                   FIEMAP_EXTENT_NOT_ALIGNED);
4460                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4461                         flags |= (FIEMAP_EXTENT_DELALLOC |
4462                                   FIEMAP_EXTENT_UNKNOWN);
4463                 } else if (fieinfo->fi_extents_max) {
4464                         u64 bytenr = em->block_start -
4465                                 (em->start - em->orig_start);
4466
4467                         disko = em->block_start + offset_in_extent;
4468
4469                         /*
4470                          * As btrfs supports shared space, this information
4471                          * can be exported to userspace tools via
4472                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4473                          * then we're just getting a count and we can skip the
4474                          * lookup stuff.
4475                          */
4476                         ret = btrfs_check_shared(NULL, root->fs_info,
4477                                                  root->objectid,
4478                                                  btrfs_ino(inode), bytenr);
4479                         if (ret < 0)
4480                                 goto out_free;
4481                         if (ret)
4482                                 flags |= FIEMAP_EXTENT_SHARED;
4483                         ret = 0;
4484                 }
4485                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4486                         flags |= FIEMAP_EXTENT_ENCODED;
4487                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4488                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4489
4490                 free_extent_map(em);
4491                 em = NULL;
4492                 if ((em_start >= last) || em_len == (u64)-1 ||
4493                    (last == (u64)-1 && isize <= em_end)) {
4494                         flags |= FIEMAP_EXTENT_LAST;
4495                         end = 1;
4496                 }
4497
4498                 /* now scan forward to see if this is really the last extent. */
4499                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4500                                            get_extent);
4501                 if (IS_ERR(em)) {
4502                         ret = PTR_ERR(em);
4503                         goto out;
4504                 }
4505                 if (!em) {
4506                         flags |= FIEMAP_EXTENT_LAST;
4507                         end = 1;
4508                 }
4509                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4510                                               em_len, flags);
4511                 if (ret) {
4512                         if (ret == 1)
4513                                 ret = 0;
4514                         goto out_free;
4515                 }
4516         }
4517 out_free:
4518         free_extent_map(em);
4519 out:
4520         btrfs_free_path(path);
4521         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4522                              &cached_state, GFP_NOFS);
4523         return ret;
4524 }
4525
4526 static void __free_extent_buffer(struct extent_buffer *eb)
4527 {
4528         btrfs_leak_debug_del(&eb->leak_list);
4529         kmem_cache_free(extent_buffer_cache, eb);
4530 }
4531
4532 int extent_buffer_under_io(struct extent_buffer *eb)
4533 {
4534         return (atomic_read(&eb->io_pages) ||
4535                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4536                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4537 }
4538
4539 /*
4540  * Helper for releasing extent buffer page.
4541  */
4542 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4543 {
4544         unsigned long index;
4545         struct page *page;
4546         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4547
4548         BUG_ON(extent_buffer_under_io(eb));
4549
4550         index = num_extent_pages(eb->start, eb->len);
4551         if (index == 0)
4552                 return;
4553
4554         do {
4555                 index--;
4556                 page = eb->pages[index];
4557                 if (!page)
4558                         continue;
4559                 if (mapped)
4560                         spin_lock(&page->mapping->private_lock);
4561                 /*
4562                  * We do this since we'll remove the pages after we've
4563                  * removed the eb from the radix tree, so we could race
4564                  * and have this page now attached to the new eb.  So
4565                  * only clear page_private if it's still connected to
4566                  * this eb.
4567                  */
4568                 if (PagePrivate(page) &&
4569                     page->private == (unsigned long)eb) {
4570                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4571                         BUG_ON(PageDirty(page));
4572                         BUG_ON(PageWriteback(page));
4573                         /*
4574                          * We need to make sure we haven't be attached
4575                          * to a new eb.
4576                          */
4577                         ClearPagePrivate(page);
4578                         set_page_private(page, 0);
4579                         /* One for the page private */
4580                         page_cache_release(page);
4581                 }
4582
4583                 if (mapped)
4584                         spin_unlock(&page->mapping->private_lock);
4585
4586                 /* One for when we alloced the page */
4587                 page_cache_release(page);
4588         } while (index != 0);
4589 }
4590
4591 /*
4592  * Helper for releasing the extent buffer.
4593  */
4594 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4595 {
4596         btrfs_release_extent_buffer_page(eb);
4597         __free_extent_buffer(eb);
4598 }
4599
4600 static struct extent_buffer *
4601 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4602                       unsigned long len)
4603 {
4604         struct extent_buffer *eb = NULL;
4605
4606         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS);
4607         if (eb == NULL)
4608                 return NULL;
4609         eb->start = start;
4610         eb->len = len;
4611         eb->fs_info = fs_info;
4612         eb->bflags = 0;
4613         rwlock_init(&eb->lock);
4614         atomic_set(&eb->write_locks, 0);
4615         atomic_set(&eb->read_locks, 0);
4616         atomic_set(&eb->blocking_readers, 0);
4617         atomic_set(&eb->blocking_writers, 0);
4618         atomic_set(&eb->spinning_readers, 0);
4619         atomic_set(&eb->spinning_writers, 0);
4620         eb->lock_nested = 0;
4621         init_waitqueue_head(&eb->write_lock_wq);
4622         init_waitqueue_head(&eb->read_lock_wq);
4623
4624         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4625
4626         spin_lock_init(&eb->refs_lock);
4627         atomic_set(&eb->refs, 1);
4628         atomic_set(&eb->io_pages, 0);
4629
4630         /*
4631          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4632          */
4633         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4634                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4635         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4636
4637         return eb;
4638 }
4639
4640 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4641 {
4642         unsigned long i;
4643         struct page *p;
4644         struct extent_buffer *new;
4645         unsigned long num_pages = num_extent_pages(src->start, src->len);
4646
4647         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4648         if (new == NULL)
4649                 return NULL;
4650
4651         for (i = 0; i < num_pages; i++) {
4652                 p = alloc_page(GFP_NOFS);
4653                 if (!p) {
4654                         btrfs_release_extent_buffer(new);
4655                         return NULL;
4656                 }
4657                 attach_extent_buffer_page(new, p);
4658                 WARN_ON(PageDirty(p));
4659                 SetPageUptodate(p);
4660                 new->pages[i] = p;
4661         }
4662
4663         copy_extent_buffer(new, src, 0, 0, src->len);
4664         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4665         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4666
4667         return new;
4668 }
4669
4670 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4671                                                 u64 start)
4672 {
4673         struct extent_buffer *eb;
4674         unsigned long len;
4675         unsigned long num_pages;
4676         unsigned long i;
4677
4678         if (!fs_info) {
4679                 /*
4680                  * Called only from tests that don't always have a fs_info
4681                  * available, but we know that nodesize is 4096
4682                  */
4683                 len = 4096;
4684         } else {
4685                 len = fs_info->tree_root->nodesize;
4686         }
4687         num_pages = num_extent_pages(0, len);
4688
4689         eb = __alloc_extent_buffer(fs_info, start, len);
4690         if (!eb)
4691                 return NULL;
4692
4693         for (i = 0; i < num_pages; i++) {
4694                 eb->pages[i] = alloc_page(GFP_NOFS);
4695                 if (!eb->pages[i])
4696                         goto err;
4697         }
4698         set_extent_buffer_uptodate(eb);
4699         btrfs_set_header_nritems(eb, 0);
4700         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4701
4702         return eb;
4703 err:
4704         for (; i > 0; i--)
4705                 __free_page(eb->pages[i - 1]);
4706         __free_extent_buffer(eb);
4707         return NULL;
4708 }
4709
4710 static void check_buffer_tree_ref(struct extent_buffer *eb)
4711 {
4712         int refs;
4713         /* the ref bit is tricky.  We have to make sure it is set
4714          * if we have the buffer dirty.   Otherwise the
4715          * code to free a buffer can end up dropping a dirty
4716          * page
4717          *
4718          * Once the ref bit is set, it won't go away while the
4719          * buffer is dirty or in writeback, and it also won't
4720          * go away while we have the reference count on the
4721          * eb bumped.
4722          *
4723          * We can't just set the ref bit without bumping the
4724          * ref on the eb because free_extent_buffer might
4725          * see the ref bit and try to clear it.  If this happens
4726          * free_extent_buffer might end up dropping our original
4727          * ref by mistake and freeing the page before we are able
4728          * to add one more ref.
4729          *
4730          * So bump the ref count first, then set the bit.  If someone
4731          * beat us to it, drop the ref we added.
4732          */
4733         refs = atomic_read(&eb->refs);
4734         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4735                 return;
4736
4737         spin_lock(&eb->refs_lock);
4738         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4739                 atomic_inc(&eb->refs);
4740         spin_unlock(&eb->refs_lock);
4741 }
4742
4743 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4744                 struct page *accessed)
4745 {
4746         unsigned long num_pages, i;
4747
4748         check_buffer_tree_ref(eb);
4749
4750         num_pages = num_extent_pages(eb->start, eb->len);
4751         for (i = 0; i < num_pages; i++) {
4752                 struct page *p = eb->pages[i];
4753
4754                 if (p != accessed)
4755                         mark_page_accessed(p);
4756         }
4757 }
4758
4759 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4760                                          u64 start)
4761 {
4762         struct extent_buffer *eb;
4763
4764         rcu_read_lock();
4765         eb = radix_tree_lookup(&fs_info->buffer_radix,
4766                                start >> PAGE_CACHE_SHIFT);
4767         if (eb && atomic_inc_not_zero(&eb->refs)) {
4768                 rcu_read_unlock();
4769                 /*
4770                  * Lock our eb's refs_lock to avoid races with
4771                  * free_extent_buffer. When we get our eb it might be flagged
4772                  * with EXTENT_BUFFER_STALE and another task running
4773                  * free_extent_buffer might have seen that flag set,
4774                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4775                  * writeback flags not set) and it's still in the tree (flag
4776                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4777                  * of decrementing the extent buffer's reference count twice.
4778                  * So here we could race and increment the eb's reference count,
4779                  * clear its stale flag, mark it as dirty and drop our reference
4780                  * before the other task finishes executing free_extent_buffer,
4781                  * which would later result in an attempt to free an extent
4782                  * buffer that is dirty.
4783                  */
4784                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4785                         spin_lock(&eb->refs_lock);
4786                         spin_unlock(&eb->refs_lock);
4787                 }
4788                 mark_extent_buffer_accessed(eb, NULL);
4789                 return eb;
4790         }
4791         rcu_read_unlock();
4792
4793         return NULL;
4794 }
4795
4796 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4797 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4798                                                u64 start)
4799 {
4800         struct extent_buffer *eb, *exists = NULL;
4801         int ret;
4802
4803         eb = find_extent_buffer(fs_info, start);
4804         if (eb)
4805                 return eb;
4806         eb = alloc_dummy_extent_buffer(fs_info, start);
4807         if (!eb)
4808                 return NULL;
4809         eb->fs_info = fs_info;
4810 again:
4811         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4812         if (ret)
4813                 goto free_eb;
4814         spin_lock(&fs_info->buffer_lock);
4815         ret = radix_tree_insert(&fs_info->buffer_radix,
4816                                 start >> PAGE_CACHE_SHIFT, eb);
4817         spin_unlock(&fs_info->buffer_lock);
4818         radix_tree_preload_end();
4819         if (ret == -EEXIST) {
4820                 exists = find_extent_buffer(fs_info, start);
4821                 if (exists)
4822                         goto free_eb;
4823                 else
4824                         goto again;
4825         }
4826         check_buffer_tree_ref(eb);
4827         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4828
4829         /*
4830          * We will free dummy extent buffer's if they come into
4831          * free_extent_buffer with a ref count of 2, but if we are using this we
4832          * want the buffers to stay in memory until we're done with them, so
4833          * bump the ref count again.
4834          */
4835         atomic_inc(&eb->refs);
4836         return eb;
4837 free_eb:
4838         btrfs_release_extent_buffer(eb);
4839         return exists;
4840 }
4841 #endif
4842
4843 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4844                                           u64 start)
4845 {
4846         unsigned long len = fs_info->tree_root->nodesize;
4847         unsigned long num_pages = num_extent_pages(start, len);
4848         unsigned long i;
4849         unsigned long index = start >> PAGE_CACHE_SHIFT;
4850         struct extent_buffer *eb;
4851         struct extent_buffer *exists = NULL;
4852         struct page *p;
4853         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4854         int uptodate = 1;
4855         int ret;
4856
4857         eb = find_extent_buffer(fs_info, start);
4858         if (eb)
4859                 return eb;
4860
4861         eb = __alloc_extent_buffer(fs_info, start, len);
4862         if (!eb)
4863                 return NULL;
4864
4865         for (i = 0; i < num_pages; i++, index++) {
4866                 p = find_or_create_page(mapping, index, GFP_NOFS);
4867                 if (!p)
4868                         goto free_eb;
4869
4870                 spin_lock(&mapping->private_lock);
4871                 if (PagePrivate(p)) {
4872                         /*
4873                          * We could have already allocated an eb for this page
4874                          * and attached one so lets see if we can get a ref on
4875                          * the existing eb, and if we can we know it's good and
4876                          * we can just return that one, else we know we can just
4877                          * overwrite page->private.
4878                          */
4879                         exists = (struct extent_buffer *)p->private;
4880                         if (atomic_inc_not_zero(&exists->refs)) {
4881                                 spin_unlock(&mapping->private_lock);
4882                                 unlock_page(p);
4883                                 page_cache_release(p);
4884                                 mark_extent_buffer_accessed(exists, p);
4885                                 goto free_eb;
4886                         }
4887                         exists = NULL;
4888
4889                         /*
4890                          * Do this so attach doesn't complain and we need to
4891                          * drop the ref the old guy had.
4892                          */
4893                         ClearPagePrivate(p);
4894                         WARN_ON(PageDirty(p));
4895                         page_cache_release(p);
4896                 }
4897                 attach_extent_buffer_page(eb, p);
4898                 spin_unlock(&mapping->private_lock);
4899                 WARN_ON(PageDirty(p));
4900                 eb->pages[i] = p;
4901                 if (!PageUptodate(p))
4902                         uptodate = 0;
4903
4904                 /*
4905                  * see below about how we avoid a nasty race with release page
4906                  * and why we unlock later
4907                  */
4908         }
4909         if (uptodate)
4910                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4911 again:
4912         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4913         if (ret)
4914                 goto free_eb;
4915
4916         spin_lock(&fs_info->buffer_lock);
4917         ret = radix_tree_insert(&fs_info->buffer_radix,
4918                                 start >> PAGE_CACHE_SHIFT, eb);
4919         spin_unlock(&fs_info->buffer_lock);
4920         radix_tree_preload_end();
4921         if (ret == -EEXIST) {
4922                 exists = find_extent_buffer(fs_info, start);
4923                 if (exists)
4924                         goto free_eb;
4925                 else
4926                         goto again;
4927         }
4928         /* add one reference for the tree */
4929         check_buffer_tree_ref(eb);
4930         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4931
4932         /*
4933          * there is a race where release page may have
4934          * tried to find this extent buffer in the radix
4935          * but failed.  It will tell the VM it is safe to
4936          * reclaim the, and it will clear the page private bit.
4937          * We must make sure to set the page private bit properly
4938          * after the extent buffer is in the radix tree so
4939          * it doesn't get lost
4940          */
4941         SetPageChecked(eb->pages[0]);
4942         for (i = 1; i < num_pages; i++) {
4943                 p = eb->pages[i];
4944                 ClearPageChecked(p);
4945                 unlock_page(p);
4946         }
4947         unlock_page(eb->pages[0]);
4948         return eb;
4949
4950 free_eb:
4951         WARN_ON(!atomic_dec_and_test(&eb->refs));
4952         for (i = 0; i < num_pages; i++) {
4953                 if (eb->pages[i])
4954                         unlock_page(eb->pages[i]);
4955         }
4956
4957         btrfs_release_extent_buffer(eb);
4958         return exists;
4959 }
4960
4961 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4962 {
4963         struct extent_buffer *eb =
4964                         container_of(head, struct extent_buffer, rcu_head);
4965
4966         __free_extent_buffer(eb);
4967 }
4968
4969 /* Expects to have eb->eb_lock already held */
4970 static int release_extent_buffer(struct extent_buffer *eb)
4971 {
4972         WARN_ON(atomic_read(&eb->refs) == 0);
4973         if (atomic_dec_and_test(&eb->refs)) {
4974                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
4975                         struct btrfs_fs_info *fs_info = eb->fs_info;
4976
4977                         spin_unlock(&eb->refs_lock);
4978
4979                         spin_lock(&fs_info->buffer_lock);
4980                         radix_tree_delete(&fs_info->buffer_radix,
4981                                           eb->start >> PAGE_CACHE_SHIFT);
4982                         spin_unlock(&fs_info->buffer_lock);
4983                 } else {
4984                         spin_unlock(&eb->refs_lock);
4985                 }
4986
4987                 /* Should be safe to release our pages at this point */
4988                 btrfs_release_extent_buffer_page(eb);
4989 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4990                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
4991                         __free_extent_buffer(eb);
4992                         return 1;
4993                 }
4994 #endif
4995                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4996                 return 1;
4997         }
4998         spin_unlock(&eb->refs_lock);
4999
5000         return 0;
5001 }
5002
5003 void free_extent_buffer(struct extent_buffer *eb)
5004 {
5005         int refs;
5006         int old;
5007         if (!eb)
5008                 return;
5009
5010         while (1) {
5011                 refs = atomic_read(&eb->refs);
5012                 if (refs <= 3)
5013                         break;
5014                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5015                 if (old == refs)
5016                         return;
5017         }
5018
5019         spin_lock(&eb->refs_lock);
5020         if (atomic_read(&eb->refs) == 2 &&
5021             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5022                 atomic_dec(&eb->refs);
5023
5024         if (atomic_read(&eb->refs) == 2 &&
5025             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5026             !extent_buffer_under_io(eb) &&
5027             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5028                 atomic_dec(&eb->refs);
5029
5030         /*
5031          * I know this is terrible, but it's temporary until we stop tracking
5032          * the uptodate bits and such for the extent buffers.
5033          */
5034         release_extent_buffer(eb);
5035 }
5036
5037 void free_extent_buffer_stale(struct extent_buffer *eb)
5038 {
5039         if (!eb)
5040                 return;
5041
5042         spin_lock(&eb->refs_lock);
5043         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5044
5045         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5046             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5047                 atomic_dec(&eb->refs);
5048         release_extent_buffer(eb);
5049 }
5050
5051 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5052 {
5053         unsigned long i;
5054         unsigned long num_pages;
5055         struct page *page;
5056
5057         num_pages = num_extent_pages(eb->start, eb->len);
5058
5059         for (i = 0; i < num_pages; i++) {
5060                 page = eb->pages[i];
5061                 if (!PageDirty(page))
5062                         continue;
5063
5064                 lock_page(page);
5065                 WARN_ON(!PagePrivate(page));
5066
5067                 clear_page_dirty_for_io(page);
5068                 spin_lock_irq(&page->mapping->tree_lock);
5069                 if (!PageDirty(page)) {
5070                         radix_tree_tag_clear(&page->mapping->page_tree,
5071                                                 page_index(page),
5072                                                 PAGECACHE_TAG_DIRTY);
5073                 }
5074                 spin_unlock_irq(&page->mapping->tree_lock);
5075                 ClearPageError(page);
5076                 unlock_page(page);
5077         }
5078         WARN_ON(atomic_read(&eb->refs) == 0);
5079 }
5080
5081 int set_extent_buffer_dirty(struct extent_buffer *eb)
5082 {
5083         unsigned long i;
5084         unsigned long num_pages;
5085         int was_dirty = 0;
5086
5087         check_buffer_tree_ref(eb);
5088
5089         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5090
5091         num_pages = num_extent_pages(eb->start, eb->len);
5092         WARN_ON(atomic_read(&eb->refs) == 0);
5093         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5094
5095         for (i = 0; i < num_pages; i++)
5096                 set_page_dirty(eb->pages[i]);
5097         return was_dirty;
5098 }
5099
5100 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
5101 {
5102         unsigned long i;
5103         struct page *page;
5104         unsigned long num_pages;
5105
5106         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5107         num_pages = num_extent_pages(eb->start, eb->len);
5108         for (i = 0; i < num_pages; i++) {
5109                 page = eb->pages[i];
5110                 if (page)
5111                         ClearPageUptodate(page);
5112         }
5113         return 0;
5114 }
5115
5116 int set_extent_buffer_uptodate(struct extent_buffer *eb)
5117 {
5118         unsigned long i;
5119         struct page *page;
5120         unsigned long num_pages;
5121
5122         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5123         num_pages = num_extent_pages(eb->start, eb->len);
5124         for (i = 0; i < num_pages; i++) {
5125                 page = eb->pages[i];
5126                 SetPageUptodate(page);
5127         }
5128         return 0;
5129 }
5130
5131 int extent_buffer_uptodate(struct extent_buffer *eb)
5132 {
5133         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5134 }
5135
5136 int read_extent_buffer_pages(struct extent_io_tree *tree,
5137                              struct extent_buffer *eb, u64 start, int wait,
5138                              get_extent_t *get_extent, int mirror_num)
5139 {
5140         unsigned long i;
5141         unsigned long start_i;
5142         struct page *page;
5143         int err;
5144         int ret = 0;
5145         int locked_pages = 0;
5146         int all_uptodate = 1;
5147         unsigned long num_pages;
5148         unsigned long num_reads = 0;
5149         struct bio *bio = NULL;
5150         unsigned long bio_flags = 0;
5151
5152         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5153                 return 0;
5154
5155         if (start) {
5156                 WARN_ON(start < eb->start);
5157                 start_i = (start >> PAGE_CACHE_SHIFT) -
5158                         (eb->start >> PAGE_CACHE_SHIFT);
5159         } else {
5160                 start_i = 0;
5161         }
5162
5163         num_pages = num_extent_pages(eb->start, eb->len);
5164         for (i = start_i; i < num_pages; i++) {
5165                 page = eb->pages[i];
5166                 if (wait == WAIT_NONE) {
5167                         if (!trylock_page(page))
5168                                 goto unlock_exit;
5169                 } else {
5170                         lock_page(page);
5171                 }
5172                 locked_pages++;
5173                 if (!PageUptodate(page)) {
5174                         num_reads++;
5175                         all_uptodate = 0;
5176                 }
5177         }
5178         if (all_uptodate) {
5179                 if (start_i == 0)
5180                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5181                 goto unlock_exit;
5182         }
5183
5184         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5185         eb->read_mirror = 0;
5186         atomic_set(&eb->io_pages, num_reads);
5187         for (i = start_i; i < num_pages; i++) {
5188                 page = eb->pages[i];
5189                 if (!PageUptodate(page)) {
5190                         ClearPageError(page);
5191                         err = __extent_read_full_page(tree, page,
5192                                                       get_extent, &bio,
5193                                                       mirror_num, &bio_flags,
5194                                                       READ | REQ_META);
5195                         if (err)
5196                                 ret = err;
5197                 } else {
5198                         unlock_page(page);
5199                 }
5200         }
5201
5202         if (bio) {
5203                 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
5204                                      bio_flags);
5205                 if (err)
5206                         return err;
5207         }
5208
5209         if (ret || wait != WAIT_COMPLETE)
5210                 return ret;
5211
5212         for (i = start_i; i < num_pages; i++) {
5213                 page = eb->pages[i];
5214                 wait_on_page_locked(page);
5215                 if (!PageUptodate(page))
5216                         ret = -EIO;
5217         }
5218
5219         return ret;
5220
5221 unlock_exit:
5222         i = start_i;
5223         while (locked_pages > 0) {
5224                 page = eb->pages[i];
5225                 i++;
5226                 unlock_page(page);
5227                 locked_pages--;
5228         }
5229         return ret;
5230 }
5231
5232 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5233                         unsigned long start,
5234                         unsigned long len)
5235 {
5236         size_t cur;
5237         size_t offset;
5238         struct page *page;
5239         char *kaddr;
5240         char *dst = (char *)dstv;
5241         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5242         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5243
5244         WARN_ON(start > eb->len);
5245         WARN_ON(start + len > eb->start + eb->len);
5246
5247         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5248
5249         while (len > 0) {
5250                 page = eb->pages[i];
5251
5252                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5253                 kaddr = page_address(page);
5254                 memcpy(dst, kaddr + offset, cur);
5255
5256                 dst += cur;
5257                 len -= cur;
5258                 offset = 0;
5259                 i++;
5260         }
5261 }
5262
5263 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5264                         unsigned long start,
5265                         unsigned long len)
5266 {
5267         size_t cur;
5268         size_t offset;
5269         struct page *page;
5270         char *kaddr;
5271         char __user *dst = (char __user *)dstv;
5272         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5273         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5274         int ret = 0;
5275
5276         WARN_ON(start > eb->len);
5277         WARN_ON(start + len > eb->start + eb->len);
5278
5279         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5280
5281         while (len > 0) {
5282                 page = eb->pages[i];
5283
5284                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5285                 kaddr = page_address(page);
5286                 if (copy_to_user(dst, kaddr + offset, cur)) {
5287                         ret = -EFAULT;
5288                         break;
5289                 }
5290
5291                 dst += cur;
5292                 len -= cur;
5293                 offset = 0;
5294                 i++;
5295         }
5296
5297         return ret;
5298 }
5299
5300 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5301                                unsigned long min_len, char **map,
5302                                unsigned long *map_start,
5303                                unsigned long *map_len)
5304 {
5305         size_t offset = start & (PAGE_CACHE_SIZE - 1);
5306         char *kaddr;
5307         struct page *p;
5308         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5309         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5310         unsigned long end_i = (start_offset + start + min_len - 1) >>
5311                 PAGE_CACHE_SHIFT;
5312
5313         if (i != end_i)
5314                 return -EINVAL;
5315
5316         if (i == 0) {
5317                 offset = start_offset;
5318                 *map_start = 0;
5319         } else {
5320                 offset = 0;
5321                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
5322         }
5323
5324         if (start + min_len > eb->len) {
5325                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
5326                        "wanted %lu %lu\n",
5327                        eb->start, eb->len, start, min_len);
5328                 return -EINVAL;
5329         }
5330
5331         p = eb->pages[i];
5332         kaddr = page_address(p);
5333         *map = kaddr + offset;
5334         *map_len = PAGE_CACHE_SIZE - offset;
5335         return 0;
5336 }
5337
5338 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5339                           unsigned long start,
5340                           unsigned long len)
5341 {
5342         size_t cur;
5343         size_t offset;
5344         struct page *page;
5345         char *kaddr;
5346         char *ptr = (char *)ptrv;
5347         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5348         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5349         int ret = 0;
5350
5351         WARN_ON(start > eb->len);
5352         WARN_ON(start + len > eb->start + eb->len);
5353
5354         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5355
5356         while (len > 0) {
5357                 page = eb->pages[i];
5358
5359                 cur = min(len, (PAGE_CACHE_SIZE - offset));
5360
5361                 kaddr = page_address(page);
5362                 ret = memcmp(ptr, kaddr + offset, cur);
5363                 if (ret)
5364                         break;
5365
5366                 ptr += cur;
5367                 len -= cur;
5368                 offset = 0;
5369                 i++;
5370         }
5371         return ret;
5372 }
5373
5374 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5375                          unsigned long start, unsigned long len)
5376 {
5377         size_t cur;
5378         size_t offset;
5379         struct page *page;
5380         char *kaddr;
5381         char *src = (char *)srcv;
5382         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5383         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5384
5385         WARN_ON(start > eb->len);
5386         WARN_ON(start + len > eb->start + eb->len);
5387
5388         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5389
5390         while (len > 0) {
5391                 page = eb->pages[i];
5392                 WARN_ON(!PageUptodate(page));
5393
5394                 cur = min(len, PAGE_CACHE_SIZE - offset);
5395                 kaddr = page_address(page);
5396                 memcpy(kaddr + offset, src, cur);
5397
5398                 src += cur;
5399                 len -= cur;
5400                 offset = 0;
5401                 i++;
5402         }
5403 }
5404
5405 void memset_extent_buffer(struct extent_buffer *eb, char c,
5406                           unsigned long start, unsigned long len)
5407 {
5408         size_t cur;
5409         size_t offset;
5410         struct page *page;
5411         char *kaddr;
5412         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5413         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5414
5415         WARN_ON(start > eb->len);
5416         WARN_ON(start + len > eb->start + eb->len);
5417
5418         offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5419
5420         while (len > 0) {
5421                 page = eb->pages[i];
5422                 WARN_ON(!PageUptodate(page));
5423
5424                 cur = min(len, PAGE_CACHE_SIZE - offset);
5425                 kaddr = page_address(page);
5426                 memset(kaddr + offset, c, cur);
5427
5428                 len -= cur;
5429                 offset = 0;
5430                 i++;
5431         }
5432 }
5433
5434 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5435                         unsigned long dst_offset, unsigned long src_offset,
5436                         unsigned long len)
5437 {
5438         u64 dst_len = dst->len;
5439         size_t cur;
5440         size_t offset;
5441         struct page *page;
5442         char *kaddr;
5443         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5444         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5445
5446         WARN_ON(src->len != dst_len);
5447
5448         offset = (start_offset + dst_offset) &
5449                 (PAGE_CACHE_SIZE - 1);
5450
5451         while (len > 0) {
5452                 page = dst->pages[i];
5453                 WARN_ON(!PageUptodate(page));
5454
5455                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5456
5457                 kaddr = page_address(page);
5458                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5459
5460                 src_offset += cur;
5461                 len -= cur;
5462                 offset = 0;
5463                 i++;
5464         }
5465 }
5466
5467 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5468 {
5469         unsigned long distance = (src > dst) ? src - dst : dst - src;
5470         return distance < len;
5471 }
5472
5473 static void copy_pages(struct page *dst_page, struct page *src_page,
5474                        unsigned long dst_off, unsigned long src_off,
5475                        unsigned long len)
5476 {
5477         char *dst_kaddr = page_address(dst_page);
5478         char *src_kaddr;
5479         int must_memmove = 0;
5480
5481         if (dst_page != src_page) {
5482                 src_kaddr = page_address(src_page);
5483         } else {
5484                 src_kaddr = dst_kaddr;
5485                 if (areas_overlap(src_off, dst_off, len))
5486                         must_memmove = 1;
5487         }
5488
5489         if (must_memmove)
5490                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5491         else
5492                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5493 }
5494
5495 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5496                            unsigned long src_offset, unsigned long len)
5497 {
5498         size_t cur;
5499         size_t dst_off_in_page;
5500         size_t src_off_in_page;
5501         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5502         unsigned long dst_i;
5503         unsigned long src_i;
5504
5505         if (src_offset + len > dst->len) {
5506                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5507                        "len %lu dst len %lu\n", src_offset, len, dst->len);
5508                 BUG_ON(1);
5509         }
5510         if (dst_offset + len > dst->len) {
5511                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5512                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
5513                 BUG_ON(1);
5514         }
5515
5516         while (len > 0) {
5517                 dst_off_in_page = (start_offset + dst_offset) &
5518                         (PAGE_CACHE_SIZE - 1);
5519                 src_off_in_page = (start_offset + src_offset) &
5520                         (PAGE_CACHE_SIZE - 1);
5521
5522                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5523                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5524
5525                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5526                                                src_off_in_page));
5527                 cur = min_t(unsigned long, cur,
5528                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5529
5530                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5531                            dst_off_in_page, src_off_in_page, cur);
5532
5533                 src_offset += cur;
5534                 dst_offset += cur;
5535                 len -= cur;
5536         }
5537 }
5538
5539 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5540                            unsigned long src_offset, unsigned long len)
5541 {
5542         size_t cur;
5543         size_t dst_off_in_page;
5544         size_t src_off_in_page;
5545         unsigned long dst_end = dst_offset + len - 1;
5546         unsigned long src_end = src_offset + len - 1;
5547         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5548         unsigned long dst_i;
5549         unsigned long src_i;
5550
5551         if (src_offset + len > dst->len) {
5552                 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5553                        "len %lu len %lu\n", src_offset, len, dst->len);
5554                 BUG_ON(1);
5555         }
5556         if (dst_offset + len > dst->len) {
5557                 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5558                        "len %lu len %lu\n", dst_offset, len, dst->len);
5559                 BUG_ON(1);
5560         }
5561         if (dst_offset < src_offset) {
5562                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5563                 return;
5564         }
5565         while (len > 0) {
5566                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5567                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5568
5569                 dst_off_in_page = (start_offset + dst_end) &
5570                         (PAGE_CACHE_SIZE - 1);
5571                 src_off_in_page = (start_offset + src_end) &
5572                         (PAGE_CACHE_SIZE - 1);
5573
5574                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5575                 cur = min(cur, dst_off_in_page + 1);
5576                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5577                            dst_off_in_page - cur + 1,
5578                            src_off_in_page - cur + 1, cur);
5579
5580                 dst_end -= cur;
5581                 src_end -= cur;
5582                 len -= cur;
5583         }
5584 }
5585
5586 int try_release_extent_buffer(struct page *page)
5587 {
5588         struct extent_buffer *eb;
5589
5590         /*
5591          * We need to make sure noboody is attaching this page to an eb right
5592          * now.
5593          */
5594         spin_lock(&page->mapping->private_lock);
5595         if (!PagePrivate(page)) {
5596                 spin_unlock(&page->mapping->private_lock);
5597                 return 1;
5598         }
5599
5600         eb = (struct extent_buffer *)page->private;
5601         BUG_ON(!eb);
5602
5603         /*
5604          * This is a little awful but should be ok, we need to make sure that
5605          * the eb doesn't disappear out from under us while we're looking at
5606          * this page.
5607          */
5608         spin_lock(&eb->refs_lock);
5609         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5610                 spin_unlock(&eb->refs_lock);
5611                 spin_unlock(&page->mapping->private_lock);
5612                 return 0;
5613         }
5614         spin_unlock(&page->mapping->private_lock);
5615
5616         /*
5617          * If tree ref isn't set then we know the ref on this eb is a real ref,
5618          * so just return, this page will likely be freed soon anyway.
5619          */
5620         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5621                 spin_unlock(&eb->refs_lock);
5622                 return 0;
5623         }
5624
5625         return release_extent_buffer(eb);
5626 }