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 #include "transaction.h"
  24
  25 static struct kmem_cache *extent_state_cache;
  26 static struct kmem_cache *extent_buffer_cache;
  27 static struct bio_set *btrfs_bioset;
  28
  29 static inline bool extent_state_in_tree(const struct extent_state *state)
  30 {
  31         return !RB_EMPTY_NODE(&state->rb_node);
  32 }
  33
  34 #ifdef CONFIG_BTRFS_DEBUG
  35 static LIST_HEAD(buffers);
  36 static LIST_HEAD(states);
  37
  38 static DEFINE_SPINLOCK(leak_lock);
  39
  40 static inline
  41 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
  42 {
  43         unsigned long flags;
  44
  45         spin_lock_irqsave(&leak_lock, flags);
  46         list_add(new, head);
  47         spin_unlock_irqrestore(&leak_lock, flags);
  48 }
  49
  50 static inline
  51 void btrfs_leak_debug_del(struct list_head *entry)
  52 {
  53         unsigned long flags;
  54
  55         spin_lock_irqsave(&leak_lock, flags);
  56         list_del(entry);
  57         spin_unlock_irqrestore(&leak_lock, flags);
  58 }
  59
  60 static inline
  61 void btrfs_leak_debug_check(void)
  62 {
  63         struct extent_state *state;
  64         struct extent_buffer *eb;
  65
  66         while (!list_empty(&states)) {
  67                 state = list_entry(states.next, struct extent_state, leak_list);
  68                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
  69                        state->start, state->end, state->state,
  70                        extent_state_in_tree(state),
  71                        atomic_read(&state->refs));
  72                 list_del(&state->leak_list);
  73                 kmem_cache_free(extent_state_cache, state);
  74         }
  75
  76         while (!list_empty(&buffers)) {
  77                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
  78                 pr_err("BTRFS: buffer leak start %llu len %lu 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                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
 100                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
 101                         caller, btrfs_ino(BTRFS_I(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 REQ_SYNC */
 131         unsigned int sync_io:1;
 132 };
 133
 134 static void add_extent_changeset(struct extent_state *state, unsigned bits,
 135                                  struct extent_changeset *changeset,
 136                                  int set)
 137 {
 138         int ret;
 139
 140         if (!changeset)
 141                 return;
 142         if (set && (state->state & bits) == bits)
 143                 return;
 144         if (!set && (state->state & bits) == 0)
 145                 return;
 146         changeset->bytes_changed += state->end - state->start + 1;
 147         ret = ulist_add(&changeset->range_changed, state->start, state->end,
 148                         GFP_ATOMIC);
 149         /* ENOMEM */
 150         BUG_ON(ret < 0);
 151 }
 152
 153 static noinline void flush_write_bio(void *data);
 154 static inline struct btrfs_fs_info *
 155 tree_fs_info(struct extent_io_tree *tree)
 156 {
 157         if (!tree->mapping)
 158                 return NULL;
 159         return btrfs_sb(tree->mapping->host->i_sb);
 160 }
 161
 162 int __init extent_io_init(void)
 163 {
 164         extent_state_cache = kmem_cache_create("btrfs_extent_state",
 165                         sizeof(struct extent_state), 0,
 166                         SLAB_MEM_SPREAD, NULL);
 167         if (!extent_state_cache)
 168                 return -ENOMEM;
 169
 170         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
 171                         sizeof(struct extent_buffer), 0,
 172                         SLAB_MEM_SPREAD, NULL);
 173         if (!extent_buffer_cache)
 174                 goto free_state_cache;
 175
 176         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
 177                                      offsetof(struct btrfs_io_bio, bio));
 178         if (!btrfs_bioset)
 179                 goto free_buffer_cache;
 180
 181         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
 182                 goto free_bioset;
 183
 184         return 0;
 185
 186 free_bioset:
 187         bioset_free(btrfs_bioset);
 188         btrfs_bioset = NULL;
 189
 190 free_buffer_cache:
 191         kmem_cache_destroy(extent_buffer_cache);
 192         extent_buffer_cache = NULL;
 193
 194 free_state_cache:
 195         kmem_cache_destroy(extent_state_cache);
 196         extent_state_cache = NULL;
 197         return -ENOMEM;
 198 }
 199
 200 void extent_io_exit(void)
 201 {
 202         btrfs_leak_debug_check();
 203
 204         /*
 205          * Make sure all delayed rcu free are flushed before we
 206          * destroy caches.
 207          */
 208         rcu_barrier();
 209         kmem_cache_destroy(extent_state_cache);
 210         kmem_cache_destroy(extent_buffer_cache);
 211         if (btrfs_bioset)
 212                 bioset_free(btrfs_bioset);
 213 }
 214
 215 void extent_io_tree_init(struct extent_io_tree *tree,
 216                          struct address_space *mapping)
 217 {
 218         tree->state = RB_ROOT;
 219         tree->ops = NULL;
 220         tree->dirty_bytes = 0;
 221         spin_lock_init(&tree->lock);
 222         tree->mapping = mapping;
 223 }
 224
 225 static struct extent_state *alloc_extent_state(gfp_t mask)
 226 {
 227         struct extent_state *state;
 228
 229         /*
 230          * The given mask might be not appropriate for the slab allocator,
 231          * drop the unsupported bits
 232          */
 233         mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
 234         state = kmem_cache_alloc(extent_state_cache, mask);
 235         if (!state)
 236                 return state;
 237         state->state = 0;
 238         state->failrec = NULL;
 239         RB_CLEAR_NODE(&state->rb_node);
 240         btrfs_leak_debug_add(&state->leak_list, &states);
 241         atomic_set(&state->refs, 1);
 242         init_waitqueue_head(&state->wq);
 243         trace_alloc_extent_state(state, mask, _RET_IP_);
 244         return state;
 245 }
 246
 247 void free_extent_state(struct extent_state *state)
 248 {
 249         if (!state)
 250                 return;
 251         if (atomic_dec_and_test(&state->refs)) {
 252                 WARN_ON(extent_state_in_tree(state));
 253                 btrfs_leak_debug_del(&state->leak_list);
 254                 trace_free_extent_state(state, _RET_IP_);
 255                 kmem_cache_free(extent_state_cache, state);
 256         }
 257 }
 258
 259 static struct rb_node *tree_insert(struct rb_root *root,
 260                                    struct rb_node *search_start,
 261                                    u64 offset,
 262                                    struct rb_node *node,
 263                                    struct rb_node ***p_in,
 264                                    struct rb_node **parent_in)
 265 {
 266         struct rb_node **p;
 267         struct rb_node *parent = NULL;
 268         struct tree_entry *entry;
 269
 270         if (p_in && parent_in) {
 271                 p = *p_in;
 272                 parent = *parent_in;
 273                 goto do_insert;
 274         }
 275
 276         p = search_start ? &search_start : &root->rb_node;
 277         while (*p) {
 278                 parent = *p;
 279                 entry = rb_entry(parent, struct tree_entry, rb_node);
 280
 281                 if (offset < entry->start)
 282                         p = &(*p)->rb_left;
 283                 else if (offset > entry->end)
 284                         p = &(*p)->rb_right;
 285                 else
 286                         return parent;
 287         }
 288
 289 do_insert:
 290         rb_link_node(node, parent, p);
 291         rb_insert_color(node, root);
 292         return NULL;
 293 }
 294
 295 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
 296                                       struct rb_node **prev_ret,
 297                                       struct rb_node **next_ret,
 298                                       struct rb_node ***p_ret,
 299                                       struct rb_node **parent_ret)
 300 {
 301         struct rb_root *root = &tree->state;
 302         struct rb_node **n = &root->rb_node;
 303         struct rb_node *prev = NULL;
 304         struct rb_node *orig_prev = NULL;
 305         struct tree_entry *entry;
 306         struct tree_entry *prev_entry = NULL;
 307
 308         while (*n) {
 309                 prev = *n;
 310                 entry = rb_entry(prev, struct tree_entry, rb_node);
 311                 prev_entry = entry;
 312
 313                 if (offset < entry->start)
 314                         n = &(*n)->rb_left;
 315                 else if (offset > entry->end)
 316                         n = &(*n)->rb_right;
 317                 else
 318                         return *n;
 319         }
 320
 321         if (p_ret)
 322                 *p_ret = n;
 323         if (parent_ret)
 324                 *parent_ret = prev;
 325
 326         if (prev_ret) {
 327                 orig_prev = prev;
 328                 while (prev && offset > prev_entry->end) {
 329                         prev = rb_next(prev);
 330                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 331                 }
 332                 *prev_ret = prev;
 333                 prev = orig_prev;
 334         }
 335
 336         if (next_ret) {
 337                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 338                 while (prev && offset < prev_entry->start) {
 339                         prev = rb_prev(prev);
 340                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
 341                 }
 342                 *next_ret = prev;
 343         }
 344         return NULL;
 345 }
 346
 347 static inline struct rb_node *
 348 tree_search_for_insert(struct extent_io_tree *tree,
 349                        u64 offset,
 350                        struct rb_node ***p_ret,
 351                        struct rb_node **parent_ret)
 352 {
 353         struct rb_node *prev = NULL;
 354         struct rb_node *ret;
 355
 356         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
 357         if (!ret)
 358                 return prev;
 359         return ret;
 360 }
 361
 362 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
 363                                           u64 offset)
 364 {
 365         return tree_search_for_insert(tree, offset, NULL, NULL);
 366 }
 367
 368 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
 369                      struct extent_state *other)
 370 {
 371         if (tree->ops && tree->ops->merge_extent_hook)
 372                 tree->ops->merge_extent_hook(tree->mapping->host, new,
 373                                              other);
 374 }
 375
 376 /*
 377  * utility function to look for merge candidates inside a given range.
 378  * Any extents with matching state are merged together into a single
 379  * extent in the tree.  Extents with EXTENT_IO in their state field
 380  * are not merged because the end_io handlers need to be able to do
 381  * operations on them without sleeping (or doing allocations/splits).
 382  *
 383  * This should be called with the tree lock held.
 384  */
 385 static void merge_state(struct extent_io_tree *tree,
 386                         struct extent_state *state)
 387 {
 388         struct extent_state *other;
 389         struct rb_node *other_node;
 390
 391         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 392                 return;
 393
 394         other_node = rb_prev(&state->rb_node);
 395         if (other_node) {
 396                 other = rb_entry(other_node, struct extent_state, rb_node);
 397                 if (other->end == state->start - 1 &&
 398                     other->state == state->state) {
 399                         merge_cb(tree, state, other);
 400                         state->start = other->start;
 401                         rb_erase(&other->rb_node, &tree->state);
 402                         RB_CLEAR_NODE(&other->rb_node);
 403                         free_extent_state(other);
 404                 }
 405         }
 406         other_node = rb_next(&state->rb_node);
 407         if (other_node) {
 408                 other = rb_entry(other_node, struct extent_state, rb_node);
 409                 if (other->start == state->end + 1 &&
 410                     other->state == state->state) {
 411                         merge_cb(tree, state, other);
 412                         state->end = other->end;
 413                         rb_erase(&other->rb_node, &tree->state);
 414                         RB_CLEAR_NODE(&other->rb_node);
 415                         free_extent_state(other);
 416                 }
 417         }
 418 }
 419
 420 static void set_state_cb(struct extent_io_tree *tree,
 421                          struct extent_state *state, unsigned *bits)
 422 {
 423         if (tree->ops && tree->ops->set_bit_hook)
 424                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
 425 }
 426
 427 static void clear_state_cb(struct extent_io_tree *tree,
 428                            struct extent_state *state, unsigned *bits)
 429 {
 430         if (tree->ops && tree->ops->clear_bit_hook)
 431                 tree->ops->clear_bit_hook(BTRFS_I(tree->mapping->host),
 432                                 state, bits);
 433 }
 434
 435 static void set_state_bits(struct extent_io_tree *tree,
 436                            struct extent_state *state, unsigned *bits,
 437                            struct extent_changeset *changeset);
 438
 439 /*
 440  * insert an extent_state struct into the tree.  'bits' are set on the
 441  * struct before it is inserted.
 442  *
 443  * This may return -EEXIST if the extent is already there, in which case the
 444  * state struct is freed.
 445  *
 446  * The tree lock is not taken internally.  This is a utility function and
 447  * probably isn't what you want to call (see set/clear_extent_bit).
 448  */
 449 static int insert_state(struct extent_io_tree *tree,
 450                         struct extent_state *state, u64 start, u64 end,
 451                         struct rb_node ***p,
 452                         struct rb_node **parent,
 453                         unsigned *bits, struct extent_changeset *changeset)
 454 {
 455         struct rb_node *node;
 456
 457         if (end < start)
 458                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
 459                        end, start);
 460         state->start = start;
 461         state->end = end;
 462
 463         set_state_bits(tree, state, bits, changeset);
 464
 465         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
 466         if (node) {
 467                 struct extent_state *found;
 468                 found = rb_entry(node, struct extent_state, rb_node);
 469                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
 470                        found->start, found->end, start, end);
 471                 return -EEXIST;
 472         }
 473         merge_state(tree, state);
 474         return 0;
 475 }
 476
 477 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
 478                      u64 split)
 479 {
 480         if (tree->ops && tree->ops->split_extent_hook)
 481                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
 482 }
 483
 484 /*
 485  * split a given extent state struct in two, inserting the preallocated
 486  * struct 'prealloc' as the newly created second half.  'split' indicates an
 487  * offset inside 'orig' where it should be split.
 488  *
 489  * Before calling,
 490  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
 491  * are two extent state structs in the tree:
 492  * prealloc: [orig->start, split - 1]
 493  * orig: [ split, orig->end ]
 494  *
 495  * The tree locks are not taken by this function. They need to be held
 496  * by the caller.
 497  */
 498 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
 499                        struct extent_state *prealloc, u64 split)
 500 {
 501         struct rb_node *node;
 502
 503         split_cb(tree, orig, split);
 504
 505         prealloc->start = orig->start;
 506         prealloc->end = split - 1;
 507         prealloc->state = orig->state;
 508         orig->start = split;
 509
 510         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
 511                            &prealloc->rb_node, NULL, NULL);
 512         if (node) {
 513                 free_extent_state(prealloc);
 514                 return -EEXIST;
 515         }
 516         return 0;
 517 }
 518
 519 static struct extent_state *next_state(struct extent_state *state)
 520 {
 521         struct rb_node *next = rb_next(&state->rb_node);
 522         if (next)
 523                 return rb_entry(next, struct extent_state, rb_node);
 524         else
 525                 return NULL;
 526 }
 527
 528 /*
 529  * utility function to clear some bits in an extent state struct.
 530  * it will optionally wake up any one waiting on this state (wake == 1).
 531  *
 532  * If no bits are set on the state struct after clearing things, the
 533  * struct is freed and removed from the tree
 534  */
 535 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
 536                                             struct extent_state *state,
 537                                             unsigned *bits, int wake,
 538                                             struct extent_changeset *changeset)
 539 {
 540         struct extent_state *next;
 541         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
 542
 543         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
 544                 u64 range = state->end - state->start + 1;
 545                 WARN_ON(range > tree->dirty_bytes);
 546                 tree->dirty_bytes -= range;
 547         }
 548         clear_state_cb(tree, state, bits);
 549         add_extent_changeset(state, bits_to_clear, changeset, 0);
 550         state->state &= ~bits_to_clear;
 551         if (wake)
 552                 wake_up(&state->wq);
 553         if (state->state == 0) {
 554                 next = next_state(state);
 555                 if (extent_state_in_tree(state)) {
 556                         rb_erase(&state->rb_node, &tree->state);
 557                         RB_CLEAR_NODE(&state->rb_node);
 558                         free_extent_state(state);
 559                 } else {
 560                         WARN_ON(1);
 561                 }
 562         } else {
 563                 merge_state(tree, state);
 564                 next = next_state(state);
 565         }
 566         return next;
 567 }
 568
 569 static struct extent_state *
 570 alloc_extent_state_atomic(struct extent_state *prealloc)
 571 {
 572         if (!prealloc)
 573                 prealloc = alloc_extent_state(GFP_ATOMIC);
 574
 575         return prealloc;
 576 }
 577
 578 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
 579 {
 580         btrfs_panic(tree_fs_info(tree), err,
 581                     "Locking error: Extent tree was modified by another thread while locked.");
 582 }
 583
 584 /*
 585  * clear some bits on a range in the tree.  This may require splitting
 586  * or inserting elements in the tree, so the gfp mask is used to
 587  * indicate which allocations or sleeping are allowed.
 588  *
 589  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
 590  * the given range from the tree regardless of state (ie for truncate).
 591  *
 592  * the range [start, end] is inclusive.
 593  *
 594  * This takes the tree lock, and returns 0 on success and < 0 on error.
 595  */
 596 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 597                               unsigned bits, int wake, int delete,
 598                               struct extent_state **cached_state,
 599                               gfp_t mask, struct extent_changeset *changeset)
 600 {
 601         struct extent_state *state;
 602         struct extent_state *cached;
 603         struct extent_state *prealloc = NULL;
 604         struct rb_node *node;
 605         u64 last_end;
 606         int err;
 607         int clear = 0;
 608
 609         btrfs_debug_check_extent_io_range(tree, start, end);
 610
 611         if (bits & EXTENT_DELALLOC)
 612                 bits |= EXTENT_NORESERVE;
 613
 614         if (delete)
 615                 bits |= ~EXTENT_CTLBITS;
 616         bits |= EXTENT_FIRST_DELALLOC;
 617
 618         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
 619                 clear = 1;
 620 again:
 621         if (!prealloc && gfpflags_allow_blocking(mask)) {
 622                 /*
 623                  * Don't care for allocation failure here because we might end
 624                  * up not needing the pre-allocated extent state at all, which
 625                  * is the case if we only have in the tree extent states that
 626                  * cover our input range and don't cover too any other range.
 627                  * If we end up needing a new extent state we allocate it later.
 628                  */
 629                 prealloc = alloc_extent_state(mask);
 630         }
 631
 632         spin_lock(&tree->lock);
 633         if (cached_state) {
 634                 cached = *cached_state;
 635
 636                 if (clear) {
 637                         *cached_state = NULL;
 638                         cached_state = NULL;
 639                 }
 640
 641                 if (cached && extent_state_in_tree(cached) &&
 642                     cached->start <= start && cached->end > start) {
 643                         if (clear)
 644                                 atomic_dec(&cached->refs);
 645                         state = cached;
 646                         goto hit_next;
 647                 }
 648                 if (clear)
 649                         free_extent_state(cached);
 650         }
 651         /*
 652          * this search will find the extents that end after
 653          * our range starts
 654          */
 655         node = tree_search(tree, start);
 656         if (!node)
 657                 goto out;
 658         state = rb_entry(node, struct extent_state, rb_node);
 659 hit_next:
 660         if (state->start > end)
 661                 goto out;
 662         WARN_ON(state->end < start);
 663         last_end = state->end;
 664
 665         /* the state doesn't have the wanted bits, go ahead */
 666         if (!(state->state & bits)) {
 667                 state = next_state(state);
 668                 goto next;
 669         }
 670
 671         /*
 672          *     | ---- desired range ---- |
 673          *  | state | or
 674          *  | ------------- state -------------- |
 675          *
 676          * We need to split the extent we found, and may flip
 677          * bits on second half.
 678          *
 679          * If the extent we found extends past our range, we
 680          * just split and search again.  It'll get split again
 681          * the next time though.
 682          *
 683          * If the extent we found is inside our range, we clear
 684          * the desired bit on it.
 685          */
 686
 687         if (state->start < start) {
 688                 prealloc = alloc_extent_state_atomic(prealloc);
 689                 BUG_ON(!prealloc);
 690                 err = split_state(tree, state, prealloc, start);
 691                 if (err)
 692                         extent_io_tree_panic(tree, err);
 693
 694                 prealloc = NULL;
 695                 if (err)
 696                         goto out;
 697                 if (state->end <= end) {
 698                         state = clear_state_bit(tree, state, &bits, wake,
 699                                                 changeset);
 700                         goto next;
 701                 }
 702                 goto search_again;
 703         }
 704         /*
 705          * | ---- desired range ---- |
 706          *                        | state |
 707          * We need to split the extent, and clear the bit
 708          * on the first half
 709          */
 710         if (state->start <= end && state->end > end) {
 711                 prealloc = alloc_extent_state_atomic(prealloc);
 712                 BUG_ON(!prealloc);
 713                 err = split_state(tree, state, prealloc, end + 1);
 714                 if (err)
 715                         extent_io_tree_panic(tree, err);
 716
 717                 if (wake)
 718                         wake_up(&state->wq);
 719
 720                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
 721
 722                 prealloc = NULL;
 723                 goto out;
 724         }
 725
 726         state = clear_state_bit(tree, state, &bits, wake, changeset);
 727 next:
 728         if (last_end == (u64)-1)
 729                 goto out;
 730         start = last_end + 1;
 731         if (start <= end && state && !need_resched())
 732                 goto hit_next;
 733
 734 search_again:
 735         if (start > end)
 736                 goto out;
 737         spin_unlock(&tree->lock);
 738         if (gfpflags_allow_blocking(mask))
 739                 cond_resched();
 740         goto again;
 741
 742 out:
 743         spin_unlock(&tree->lock);
 744         if (prealloc)
 745                 free_extent_state(prealloc);
 746
 747         return 0;
 748
 749 }
 750
 751 static void wait_on_state(struct extent_io_tree *tree,
 752                           struct extent_state *state)
 753                 __releases(tree->lock)
 754                 __acquires(tree->lock)
 755 {
 756         DEFINE_WAIT(wait);
 757         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
 758         spin_unlock(&tree->lock);
 759         schedule();
 760         spin_lock(&tree->lock);
 761         finish_wait(&state->wq, &wait);
 762 }
 763
 764 /*
 765  * waits for one or more bits to clear on a range in the state tree.
 766  * The range [start, end] is inclusive.
 767  * The tree lock is taken by this function
 768  */
 769 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 770                             unsigned long bits)
 771 {
 772         struct extent_state *state;
 773         struct rb_node *node;
 774
 775         btrfs_debug_check_extent_io_range(tree, start, end);
 776
 777         spin_lock(&tree->lock);
 778 again:
 779         while (1) {
 780                 /*
 781                  * this search will find all the extents that end after
 782                  * our range starts
 783                  */
 784                 node = tree_search(tree, start);
 785 process_node:
 786                 if (!node)
 787                         break;
 788
 789                 state = rb_entry(node, struct extent_state, rb_node);
 790
 791                 if (state->start > end)
 792                         goto out;
 793
 794                 if (state->state & bits) {
 795                         start = state->start;
 796                         atomic_inc(&state->refs);
 797                         wait_on_state(tree, state);
 798                         free_extent_state(state);
 799                         goto again;
 800                 }
 801                 start = state->end + 1;
 802
 803                 if (start > end)
 804                         break;
 805
 806                 if (!cond_resched_lock(&tree->lock)) {
 807                         node = rb_next(node);
 808                         goto process_node;
 809                 }
 810         }
 811 out:
 812         spin_unlock(&tree->lock);
 813 }
 814
 815 static void set_state_bits(struct extent_io_tree *tree,
 816                            struct extent_state *state,
 817                            unsigned *bits, struct extent_changeset *changeset)
 818 {
 819         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
 820
 821         set_state_cb(tree, state, bits);
 822         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
 823                 u64 range = state->end - state->start + 1;
 824                 tree->dirty_bytes += range;
 825         }
 826         add_extent_changeset(state, bits_to_set, changeset, 1);
 827         state->state |= bits_to_set;
 828 }
 829
 830 static void cache_state_if_flags(struct extent_state *state,
 831                                  struct extent_state **cached_ptr,
 832                                  unsigned flags)
 833 {
 834         if (cached_ptr && !(*cached_ptr)) {
 835                 if (!flags || (state->state & flags)) {
 836                         *cached_ptr = state;
 837                         atomic_inc(&state->refs);
 838                 }
 839         }
 840 }
 841
 842 static void cache_state(struct extent_state *state,
 843                         struct extent_state **cached_ptr)
 844 {
 845         return cache_state_if_flags(state, cached_ptr,
 846                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
 847 }
 848
 849 /*
 850  * set some bits on a range in the tree.  This may require allocations or
 851  * sleeping, so the gfp mask is used to indicate what is allowed.
 852  *
 853  * If any of the exclusive bits are set, this will fail with -EEXIST if some
 854  * part of the range already has the desired bits set.  The start of the
 855  * existing range is returned in failed_start in this case.
 856  *
 857  * [start, end] is inclusive This takes the tree lock.
 858  */
 859
 860 static int __must_check
 861 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
 862                  unsigned bits, unsigned exclusive_bits,
 863                  u64 *failed_start, struct extent_state **cached_state,
 864                  gfp_t mask, struct extent_changeset *changeset)
 865 {
 866         struct extent_state *state;
 867         struct extent_state *prealloc = NULL;
 868         struct rb_node *node;
 869         struct rb_node **p;
 870         struct rb_node *parent;
 871         int err = 0;
 872         u64 last_start;
 873         u64 last_end;
 874
 875         btrfs_debug_check_extent_io_range(tree, start, end);
 876
 877         bits |= EXTENT_FIRST_DELALLOC;
 878 again:
 879         if (!prealloc && gfpflags_allow_blocking(mask)) {
 880                 /*
 881                  * Don't care for allocation failure here because we might end
 882                  * up not needing the pre-allocated extent state at all, which
 883                  * is the case if we only have in the tree extent states that
 884                  * cover our input range and don't cover too any other range.
 885                  * If we end up needing a new extent state we allocate it later.
 886                  */
 887                 prealloc = alloc_extent_state(mask);
 888         }
 889
 890         spin_lock(&tree->lock);
 891         if (cached_state && *cached_state) {
 892                 state = *cached_state;
 893                 if (state->start <= start && state->end > start &&
 894                     extent_state_in_tree(state)) {
 895                         node = &state->rb_node;
 896                         goto hit_next;
 897                 }
 898         }
 899         /*
 900          * this search will find all the extents that end after
 901          * our range starts.
 902          */
 903         node = tree_search_for_insert(tree, start, &p, &parent);
 904         if (!node) {
 905                 prealloc = alloc_extent_state_atomic(prealloc);
 906                 BUG_ON(!prealloc);
 907                 err = insert_state(tree, prealloc, start, end,
 908                                    &p, &parent, &bits, changeset);
 909                 if (err)
 910                         extent_io_tree_panic(tree, err);
 911
 912                 cache_state(prealloc, cached_state);
 913                 prealloc = NULL;
 914                 goto out;
 915         }
 916         state = rb_entry(node, struct extent_state, rb_node);
 917 hit_next:
 918         last_start = state->start;
 919         last_end = state->end;
 920
 921         /*
 922          * | ---- desired range ---- |
 923          * | state |
 924          *
 925          * Just lock what we found and keep going
 926          */
 927         if (state->start == start && state->end <= end) {
 928                 if (state->state & exclusive_bits) {
 929                         *failed_start = state->start;
 930                         err = -EEXIST;
 931                         goto out;
 932                 }
 933
 934                 set_state_bits(tree, state, &bits, changeset);
 935                 cache_state(state, cached_state);
 936                 merge_state(tree, state);
 937                 if (last_end == (u64)-1)
 938                         goto out;
 939                 start = last_end + 1;
 940                 state = next_state(state);
 941                 if (start < end && state && state->start == start &&
 942                     !need_resched())
 943                         goto hit_next;
 944                 goto search_again;
 945         }
 946
 947         /*
 948          *     | ---- desired range ---- |
 949          * | state |
 950          *   or
 951          * | ------------- state -------------- |
 952          *
 953          * We need to split the extent we found, and may flip bits on
 954          * second half.
 955          *
 956          * If the extent we found extends past our
 957          * range, we just split and search again.  It'll get split
 958          * again the next time though.
 959          *
 960          * If the extent we found is inside our range, we set the
 961          * desired bit on it.
 962          */
 963         if (state->start < start) {
 964                 if (state->state & exclusive_bits) {
 965                         *failed_start = start;
 966                         err = -EEXIST;
 967                         goto out;
 968                 }
 969
 970                 prealloc = alloc_extent_state_atomic(prealloc);
 971                 BUG_ON(!prealloc);
 972                 err = split_state(tree, state, prealloc, start);
 973                 if (err)
 974                         extent_io_tree_panic(tree, err);
 975
 976                 prealloc = NULL;
 977                 if (err)
 978                         goto out;
 979                 if (state->end <= end) {
 980                         set_state_bits(tree, state, &bits, changeset);
 981                         cache_state(state, cached_state);
 982                         merge_state(tree, state);
 983                         if (last_end == (u64)-1)
 984                                 goto out;
 985                         start = last_end + 1;
 986                         state = next_state(state);
 987                         if (start < end && state && state->start == start &&
 988                             !need_resched())
 989                                 goto hit_next;
 990                 }
 991                 goto search_again;
 992         }
 993         /*
 994          * | ---- desired range ---- |
 995          *     | state | or               | state |
 996          *
 997          * There's a hole, we need to insert something in it and
 998          * ignore the extent we found.
 999          */
1000         if (state->start > start) {
1001                 u64 this_end;
1002                 if (end < last_start)
1003                         this_end = end;
1004                 else
1005                         this_end = last_start - 1;
1006
1007                 prealloc = alloc_extent_state_atomic(prealloc);
1008                 BUG_ON(!prealloc);
1009
1010                 /*
1011                  * Avoid to free 'prealloc' if it can be merged with
1012                  * the later extent.
1013                  */
1014                 err = insert_state(tree, prealloc, start, this_end,
1015                                    NULL, NULL, &bits, changeset);
1016                 if (err)
1017                         extent_io_tree_panic(tree, err);
1018
1019                 cache_state(prealloc, cached_state);
1020                 prealloc = NULL;
1021                 start = this_end + 1;
1022                 goto search_again;
1023         }
1024         /*
1025          * | ---- desired range ---- |
1026          *                        | state |
1027          * We need to split the extent, and set the bit
1028          * on the first half
1029          */
1030         if (state->start <= end && state->end > end) {
1031                 if (state->state & exclusive_bits) {
1032                         *failed_start = start;
1033                         err = -EEXIST;
1034                         goto out;
1035                 }
1036
1037                 prealloc = alloc_extent_state_atomic(prealloc);
1038                 BUG_ON(!prealloc);
1039                 err = split_state(tree, state, prealloc, end + 1);
1040                 if (err)
1041                         extent_io_tree_panic(tree, err);
1042
1043                 set_state_bits(tree, prealloc, &bits, changeset);
1044                 cache_state(prealloc, cached_state);
1045                 merge_state(tree, prealloc);
1046                 prealloc = NULL;
1047                 goto out;
1048         }
1049
1050 search_again:
1051         if (start > end)
1052                 goto out;
1053         spin_unlock(&tree->lock);
1054         if (gfpflags_allow_blocking(mask))
1055                 cond_resched();
1056         goto again;
1057
1058 out:
1059         spin_unlock(&tree->lock);
1060         if (prealloc)
1061                 free_extent_state(prealloc);
1062
1063         return err;
1064
1065 }
1066
1067 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1068                    unsigned bits, u64 * failed_start,
1069                    struct extent_state **cached_state, gfp_t mask)
1070 {
1071         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1072                                 cached_state, mask, NULL);
1073 }
1074
1075
1076 /**
1077  * convert_extent_bit - convert all bits in a given range from one bit to
1078  *                      another
1079  * @tree:       the io tree to search
1080  * @start:      the start offset in bytes
1081  * @end:        the end offset in bytes (inclusive)
1082  * @bits:       the bits to set in this range
1083  * @clear_bits: the bits to clear in this range
1084  * @cached_state:       state that we're going to cache
1085  *
1086  * This will go through and set bits for the given range.  If any states exist
1087  * already in this range they are set with the given bit and cleared of the
1088  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1089  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1090  * boundary bits like LOCK.
1091  *
1092  * All allocations are done with GFP_NOFS.
1093  */
1094 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1095                        unsigned bits, unsigned clear_bits,
1096                        struct extent_state **cached_state)
1097 {
1098         struct extent_state *state;
1099         struct extent_state *prealloc = NULL;
1100         struct rb_node *node;
1101         struct rb_node **p;
1102         struct rb_node *parent;
1103         int err = 0;
1104         u64 last_start;
1105         u64 last_end;
1106         bool first_iteration = true;
1107
1108         btrfs_debug_check_extent_io_range(tree, start, end);
1109
1110 again:
1111         if (!prealloc) {
1112                 /*
1113                  * Best effort, don't worry if extent state allocation fails
1114                  * here for the first iteration. We might have a cached state
1115                  * that matches exactly the target range, in which case no
1116                  * extent state allocations are needed. We'll only know this
1117                  * after locking the tree.
1118                  */
1119                 prealloc = alloc_extent_state(GFP_NOFS);
1120                 if (!prealloc && !first_iteration)
1121                         return -ENOMEM;
1122         }
1123
1124         spin_lock(&tree->lock);
1125         if (cached_state && *cached_state) {
1126                 state = *cached_state;
1127                 if (state->start <= start && state->end > start &&
1128                     extent_state_in_tree(state)) {
1129                         node = &state->rb_node;
1130                         goto hit_next;
1131                 }
1132         }
1133
1134         /*
1135          * this search will find all the extents that end after
1136          * our range starts.
1137          */
1138         node = tree_search_for_insert(tree, start, &p, &parent);
1139         if (!node) {
1140                 prealloc = alloc_extent_state_atomic(prealloc);
1141                 if (!prealloc) {
1142                         err = -ENOMEM;
1143                         goto out;
1144                 }
1145                 err = insert_state(tree, prealloc, start, end,
1146                                    &p, &parent, &bits, NULL);
1147                 if (err)
1148                         extent_io_tree_panic(tree, err);
1149                 cache_state(prealloc, cached_state);
1150                 prealloc = NULL;
1151                 goto out;
1152         }
1153         state = rb_entry(node, struct extent_state, rb_node);
1154 hit_next:
1155         last_start = state->start;
1156         last_end = state->end;
1157
1158         /*
1159          * | ---- desired range ---- |
1160          * | state |
1161          *
1162          * Just lock what we found and keep going
1163          */
1164         if (state->start == start && state->end <= end) {
1165                 set_state_bits(tree, state, &bits, NULL);
1166                 cache_state(state, cached_state);
1167                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1168                 if (last_end == (u64)-1)
1169                         goto out;
1170                 start = last_end + 1;
1171                 if (start < end && state && state->start == start &&
1172                     !need_resched())
1173                         goto hit_next;
1174                 goto search_again;
1175         }
1176
1177         /*
1178          *     | ---- desired range ---- |
1179          * | state |
1180          *   or
1181          * | ------------- state -------------- |
1182          *
1183          * We need to split the extent we found, and may flip bits on
1184          * second half.
1185          *
1186          * If the extent we found extends past our
1187          * range, we just split and search again.  It'll get split
1188          * again the next time though.
1189          *
1190          * If the extent we found is inside our range, we set the
1191          * desired bit on it.
1192          */
1193         if (state->start < start) {
1194                 prealloc = alloc_extent_state_atomic(prealloc);
1195                 if (!prealloc) {
1196                         err = -ENOMEM;
1197                         goto out;
1198                 }
1199                 err = split_state(tree, state, prealloc, start);
1200                 if (err)
1201                         extent_io_tree_panic(tree, err);
1202                 prealloc = NULL;
1203                 if (err)
1204                         goto out;
1205                 if (state->end <= end) {
1206                         set_state_bits(tree, state, &bits, NULL);
1207                         cache_state(state, cached_state);
1208                         state = clear_state_bit(tree, state, &clear_bits, 0,
1209                                                 NULL);
1210                         if (last_end == (u64)-1)
1211                                 goto out;
1212                         start = last_end + 1;
1213                         if (start < end && state && state->start == start &&
1214                             !need_resched())
1215                                 goto hit_next;
1216                 }
1217                 goto search_again;
1218         }
1219         /*
1220          * | ---- desired range ---- |
1221          *     | state | or               | state |
1222          *
1223          * There's a hole, we need to insert something in it and
1224          * ignore the extent we found.
1225          */
1226         if (state->start > start) {
1227                 u64 this_end;
1228                 if (end < last_start)
1229                         this_end = end;
1230                 else
1231                         this_end = last_start - 1;
1232
1233                 prealloc = alloc_extent_state_atomic(prealloc);
1234                 if (!prealloc) {
1235                         err = -ENOMEM;
1236                         goto out;
1237                 }
1238
1239                 /*
1240                  * Avoid to free 'prealloc' if it can be merged with
1241                  * the later extent.
1242                  */
1243                 err = insert_state(tree, prealloc, start, this_end,
1244                                    NULL, NULL, &bits, NULL);
1245                 if (err)
1246                         extent_io_tree_panic(tree, err);
1247                 cache_state(prealloc, cached_state);
1248                 prealloc = NULL;
1249                 start = this_end + 1;
1250                 goto search_again;
1251         }
1252         /*
1253          * | ---- desired range ---- |
1254          *                        | state |
1255          * We need to split the extent, and set the bit
1256          * on the first half
1257          */
1258         if (state->start <= end && state->end > end) {
1259                 prealloc = alloc_extent_state_atomic(prealloc);
1260                 if (!prealloc) {
1261                         err = -ENOMEM;
1262                         goto out;
1263                 }
1264
1265                 err = split_state(tree, state, prealloc, end + 1);
1266                 if (err)
1267                         extent_io_tree_panic(tree, err);
1268
1269                 set_state_bits(tree, prealloc, &bits, NULL);
1270                 cache_state(prealloc, cached_state);
1271                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1272                 prealloc = NULL;
1273                 goto out;
1274         }
1275
1276 search_again:
1277         if (start > end)
1278                 goto out;
1279         spin_unlock(&tree->lock);
1280         cond_resched();
1281         first_iteration = false;
1282         goto again;
1283
1284 out:
1285         spin_unlock(&tree->lock);
1286         if (prealloc)
1287                 free_extent_state(prealloc);
1288
1289         return err;
1290 }
1291
1292 /* wrappers around set/clear extent bit */
1293 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1294                            unsigned bits, struct extent_changeset *changeset)
1295 {
1296         /*
1297          * We don't support EXTENT_LOCKED yet, as current changeset will
1298          * record any bits changed, so for EXTENT_LOCKED case, it will
1299          * either fail with -EEXIST or changeset will record the whole
1300          * range.
1301          */
1302         BUG_ON(bits & EXTENT_LOCKED);
1303
1304         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1305                                 changeset);
1306 }
1307
1308 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1309                      unsigned bits, int wake, int delete,
1310                      struct extent_state **cached, gfp_t mask)
1311 {
1312         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1313                                   cached, mask, NULL);
1314 }
1315
1316 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1317                 unsigned bits, struct extent_changeset *changeset)
1318 {
1319         /*
1320          * Don't support EXTENT_LOCKED case, same reason as
1321          * set_record_extent_bits().
1322          */
1323         BUG_ON(bits & EXTENT_LOCKED);
1324
1325         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1326                                   changeset);
1327 }
1328
1329 /*
1330  * either insert or lock state struct between start and end use mask to tell
1331  * us if waiting is desired.
1332  */
1333 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1334                      struct extent_state **cached_state)
1335 {
1336         int err;
1337         u64 failed_start;
1338
1339         while (1) {
1340                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1341                                        EXTENT_LOCKED, &failed_start,
1342                                        cached_state, GFP_NOFS, NULL);
1343                 if (err == -EEXIST) {
1344                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1345                         start = failed_start;
1346                 } else
1347                         break;
1348                 WARN_ON(start > end);
1349         }
1350         return err;
1351 }
1352
1353 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1354 {
1355         int err;
1356         u64 failed_start;
1357
1358         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1359                                &failed_start, NULL, GFP_NOFS, NULL);
1360         if (err == -EEXIST) {
1361                 if (failed_start > start)
1362                         clear_extent_bit(tree, start, failed_start - 1,
1363                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1364                 return 0;
1365         }
1366         return 1;
1367 }
1368
1369 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1370 {
1371         unsigned long index = start >> PAGE_SHIFT;
1372         unsigned long end_index = end >> PAGE_SHIFT;
1373         struct page *page;
1374
1375         while (index <= end_index) {
1376                 page = find_get_page(inode->i_mapping, index);
1377                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1378                 clear_page_dirty_for_io(page);
1379                 put_page(page);
1380                 index++;
1381         }
1382 }
1383
1384 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1385 {
1386         unsigned long index = start >> PAGE_SHIFT;
1387         unsigned long end_index = end >> PAGE_SHIFT;
1388         struct page *page;
1389
1390         while (index <= end_index) {
1391                 page = find_get_page(inode->i_mapping, index);
1392                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1393                 __set_page_dirty_nobuffers(page);
1394                 account_page_redirty(page);
1395                 put_page(page);
1396                 index++;
1397         }
1398 }
1399
1400 /*
1401  * helper function to set both pages and extents in the tree writeback
1402  */
1403 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1404 {
1405         unsigned long index = start >> PAGE_SHIFT;
1406         unsigned long end_index = end >> PAGE_SHIFT;
1407         struct page *page;
1408
1409         while (index <= end_index) {
1410                 page = find_get_page(tree->mapping, index);
1411                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1412                 set_page_writeback(page);
1413                 put_page(page);
1414                 index++;
1415         }
1416 }
1417
1418 /* find the first state struct with 'bits' set after 'start', and
1419  * return it.  tree->lock must be held.  NULL will returned if
1420  * nothing was found after 'start'
1421  */
1422 static struct extent_state *
1423 find_first_extent_bit_state(struct extent_io_tree *tree,
1424                             u64 start, unsigned bits)
1425 {
1426         struct rb_node *node;
1427         struct extent_state *state;
1428
1429         /*
1430          * this search will find all the extents that end after
1431          * our range starts.
1432          */
1433         node = tree_search(tree, start);
1434         if (!node)
1435                 goto out;
1436
1437         while (1) {
1438                 state = rb_entry(node, struct extent_state, rb_node);
1439                 if (state->end >= start && (state->state & bits))
1440                         return state;
1441
1442                 node = rb_next(node);
1443                 if (!node)
1444                         break;
1445         }
1446 out:
1447         return NULL;
1448 }
1449
1450 /*
1451  * find the first offset in the io tree with 'bits' set. zero is
1452  * returned if we find something, and *start_ret and *end_ret are
1453  * set to reflect the state struct that was found.
1454  *
1455  * If nothing was found, 1 is returned. If found something, return 0.
1456  */
1457 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1458                           u64 *start_ret, u64 *end_ret, unsigned bits,
1459                           struct extent_state **cached_state)
1460 {
1461         struct extent_state *state;
1462         struct rb_node *n;
1463         int ret = 1;
1464
1465         spin_lock(&tree->lock);
1466         if (cached_state && *cached_state) {
1467                 state = *cached_state;
1468                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1469                         n = rb_next(&state->rb_node);
1470                         while (n) {
1471                                 state = rb_entry(n, struct extent_state,
1472                                                  rb_node);
1473                                 if (state->state & bits)
1474                                         goto got_it;
1475                                 n = rb_next(n);
1476                         }
1477                         free_extent_state(*cached_state);
1478                         *cached_state = NULL;
1479                         goto out;
1480                 }
1481                 free_extent_state(*cached_state);
1482                 *cached_state = NULL;
1483         }
1484
1485         state = find_first_extent_bit_state(tree, start, bits);
1486 got_it:
1487         if (state) {
1488                 cache_state_if_flags(state, cached_state, 0);
1489                 *start_ret = state->start;
1490                 *end_ret = state->end;
1491                 ret = 0;
1492         }
1493 out:
1494         spin_unlock(&tree->lock);
1495         return ret;
1496 }
1497
1498 /*
1499  * find a contiguous range of bytes in the file marked as delalloc, not
1500  * more than 'max_bytes'.  start and end are used to return the range,
1501  *
1502  * 1 is returned if we find something, 0 if nothing was in the tree
1503  */
1504 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1505                                         u64 *start, u64 *end, u64 max_bytes,
1506                                         struct extent_state **cached_state)
1507 {
1508         struct rb_node *node;
1509         struct extent_state *state;
1510         u64 cur_start = *start;
1511         u64 found = 0;
1512         u64 total_bytes = 0;
1513
1514         spin_lock(&tree->lock);
1515
1516         /*
1517          * this search will find all the extents that end after
1518          * our range starts.
1519          */
1520         node = tree_search(tree, cur_start);
1521         if (!node) {
1522                 if (!found)
1523                         *end = (u64)-1;
1524                 goto out;
1525         }
1526
1527         while (1) {
1528                 state = rb_entry(node, struct extent_state, rb_node);
1529                 if (found && (state->start != cur_start ||
1530                               (state->state & EXTENT_BOUNDARY))) {
1531                         goto out;
1532                 }
1533                 if (!(state->state & EXTENT_DELALLOC)) {
1534                         if (!found)
1535                                 *end = state->end;
1536                         goto out;
1537                 }
1538                 if (!found) {
1539                         *start = state->start;
1540                         *cached_state = state;
1541                         atomic_inc(&state->refs);
1542                 }
1543                 found++;
1544                 *end = state->end;
1545                 cur_start = state->end + 1;
1546                 node = rb_next(node);
1547                 total_bytes += state->end - state->start + 1;
1548                 if (total_bytes >= max_bytes)
1549                         break;
1550                 if (!node)
1551                         break;
1552         }
1553 out:
1554         spin_unlock(&tree->lock);
1555         return found;
1556 }
1557
1558 static int __process_pages_contig(struct address_space *mapping,
1559                                   struct page *locked_page,
1560                                   pgoff_t start_index, pgoff_t end_index,
1561                                   unsigned long page_ops, pgoff_t *index_ret);
1562
1563 static noinline void __unlock_for_delalloc(struct inode *inode,
1564                                            struct page *locked_page,
1565                                            u64 start, u64 end)
1566 {
1567         unsigned long index = start >> PAGE_SHIFT;
1568         unsigned long end_index = end >> PAGE_SHIFT;
1569
1570         ASSERT(locked_page);
1571         if (index == locked_page->index && end_index == index)
1572                 return;
1573
1574         __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1575                                PAGE_UNLOCK, NULL);
1576 }
1577
1578 static noinline int lock_delalloc_pages(struct inode *inode,
1579                                         struct page *locked_page,
1580                                         u64 delalloc_start,
1581                                         u64 delalloc_end)
1582 {
1583         unsigned long index = delalloc_start >> PAGE_SHIFT;
1584         unsigned long index_ret = index;
1585         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1586         int ret;
1587
1588         ASSERT(locked_page);
1589         if (index == locked_page->index && index == end_index)
1590                 return 0;
1591
1592         ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1593                                      end_index, PAGE_LOCK, &index_ret);
1594         if (ret == -EAGAIN)
1595                 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1596                                       (u64)index_ret << PAGE_SHIFT);
1597         return ret;
1598 }
1599
1600 /*
1601  * find a contiguous range of bytes in the file marked as delalloc, not
1602  * more than 'max_bytes'.  start and end are used to return the range,
1603  *
1604  * 1 is returned if we find something, 0 if nothing was in the tree
1605  */
1606 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1607                                     struct extent_io_tree *tree,
1608                                     struct page *locked_page, u64 *start,
1609                                     u64 *end, u64 max_bytes)
1610 {
1611         u64 delalloc_start;
1612         u64 delalloc_end;
1613         u64 found;
1614         struct extent_state *cached_state = NULL;
1615         int ret;
1616         int loops = 0;
1617
1618 again:
1619         /* step one, find a bunch of delalloc bytes starting at start */
1620         delalloc_start = *start;
1621         delalloc_end = 0;
1622         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1623                                     max_bytes, &cached_state);
1624         if (!found || delalloc_end <= *start) {
1625                 *start = delalloc_start;
1626                 *end = delalloc_end;
1627                 free_extent_state(cached_state);
1628                 return 0;
1629         }
1630
1631         /*
1632          * start comes from the offset of locked_page.  We have to lock
1633          * pages in order, so we can't process delalloc bytes before
1634          * locked_page
1635          */
1636         if (delalloc_start < *start)
1637                 delalloc_start = *start;
1638
1639         /*
1640          * make sure to limit the number of pages we try to lock down
1641          */
1642         if (delalloc_end + 1 - delalloc_start > max_bytes)
1643                 delalloc_end = delalloc_start + max_bytes - 1;
1644
1645         /* step two, lock all the pages after the page that has start */
1646         ret = lock_delalloc_pages(inode, locked_page,
1647                                   delalloc_start, delalloc_end);
1648         if (ret == -EAGAIN) {
1649                 /* some of the pages are gone, lets avoid looping by
1650                  * shortening the size of the delalloc range we're searching
1651                  */
1652                 free_extent_state(cached_state);
1653                 cached_state = NULL;
1654                 if (!loops) {
1655                         max_bytes = PAGE_SIZE;
1656                         loops = 1;
1657                         goto again;
1658                 } else {
1659                         found = 0;
1660                         goto out_failed;
1661                 }
1662         }
1663         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1664
1665         /* step three, lock the state bits for the whole range */
1666         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1667
1668         /* then test to make sure it is all still delalloc */
1669         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1670                              EXTENT_DELALLOC, 1, cached_state);
1671         if (!ret) {
1672                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1673                                      &cached_state, GFP_NOFS);
1674                 __unlock_for_delalloc(inode, locked_page,
1675                               delalloc_start, delalloc_end);
1676                 cond_resched();
1677                 goto again;
1678         }
1679         free_extent_state(cached_state);
1680         *start = delalloc_start;
1681         *end = delalloc_end;
1682 out_failed:
1683         return found;
1684 }
1685
1686 static int __process_pages_contig(struct address_space *mapping,
1687                                   struct page *locked_page,
1688                                   pgoff_t start_index, pgoff_t end_index,
1689                                   unsigned long page_ops, pgoff_t *index_ret)
1690 {
1691         unsigned long nr_pages = end_index - start_index + 1;
1692         unsigned long pages_locked = 0;
1693         pgoff_t index = start_index;
1694         struct page *pages[16];
1695         unsigned ret;
1696         int err = 0;
1697         int i;
1698
1699         if (page_ops & PAGE_LOCK) {
1700                 ASSERT(page_ops == PAGE_LOCK);
1701                 ASSERT(index_ret && *index_ret == start_index);
1702         }
1703
1704         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1705                 mapping_set_error(mapping, -EIO);
1706
1707         while (nr_pages > 0) {
1708                 ret = find_get_pages_contig(mapping, index,
1709                                      min_t(unsigned long,
1710                                      nr_pages, ARRAY_SIZE(pages)), pages);
1711                 if (ret == 0) {
1712                         /*
1713                          * Only if we're going to lock these pages,
1714                          * can we find nothing at @index.
1715                          */
1716                         ASSERT(page_ops & PAGE_LOCK);
1717                         err = -EAGAIN;
1718                         goto out;
1719                 }
1720
1721                 for (i = 0; i < ret; i++) {
1722                         if (page_ops & PAGE_SET_PRIVATE2)
1723                                 SetPagePrivate2(pages[i]);
1724
1725                         if (pages[i] == locked_page) {
1726                                 put_page(pages[i]);
1727                                 pages_locked++;
1728                                 continue;
1729                         }
1730                         if (page_ops & PAGE_CLEAR_DIRTY)
1731                                 clear_page_dirty_for_io(pages[i]);
1732                         if (page_ops & PAGE_SET_WRITEBACK)
1733                                 set_page_writeback(pages[i]);
1734                         if (page_ops & PAGE_SET_ERROR)
1735                                 SetPageError(pages[i]);
1736                         if (page_ops & PAGE_END_WRITEBACK)
1737                                 end_page_writeback(pages[i]);
1738                         if (page_ops & PAGE_UNLOCK)
1739                                 unlock_page(pages[i]);
1740                         if (page_ops & PAGE_LOCK) {
1741                                 lock_page(pages[i]);
1742                                 if (!PageDirty(pages[i]) ||
1743                                     pages[i]->mapping != mapping) {
1744                                         unlock_page(pages[i]);
1745                                         put_page(pages[i]);
1746                                         err = -EAGAIN;
1747                                         goto out;
1748                                 }
1749                         }
1750                         put_page(pages[i]);
1751                         pages_locked++;
1752                 }
1753                 nr_pages -= ret;
1754                 index += ret;
1755                 cond_resched();
1756         }
1757 out:
1758         if (err && index_ret)
1759                 *index_ret = start_index + pages_locked - 1;
1760         return err;
1761 }
1762
1763 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1764                                  u64 delalloc_end, struct page *locked_page,
1765                                  unsigned clear_bits,
1766                                  unsigned long page_ops)
1767 {
1768         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1769                          NULL, GFP_NOFS);
1770
1771         __process_pages_contig(inode->i_mapping, locked_page,
1772                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1773                                page_ops, NULL);
1774 }
1775
1776 /*
1777  * count the number of bytes in the tree that have a given bit(s)
1778  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1779  * cached.  The total number found is returned.
1780  */
1781 u64 count_range_bits(struct extent_io_tree *tree,
1782                      u64 *start, u64 search_end, u64 max_bytes,
1783                      unsigned bits, int contig)
1784 {
1785         struct rb_node *node;
1786         struct extent_state *state;
1787         u64 cur_start = *start;
1788         u64 total_bytes = 0;
1789         u64 last = 0;
1790         int found = 0;
1791
1792         if (WARN_ON(search_end <= cur_start))
1793                 return 0;
1794
1795         spin_lock(&tree->lock);
1796         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1797                 total_bytes = tree->dirty_bytes;
1798                 goto out;
1799         }
1800         /*
1801          * this search will find all the extents that end after
1802          * our range starts.
1803          */
1804         node = tree_search(tree, cur_start);
1805         if (!node)
1806                 goto out;
1807
1808         while (1) {
1809                 state = rb_entry(node, struct extent_state, rb_node);
1810                 if (state->start > search_end)
1811                         break;
1812                 if (contig && found && state->start > last + 1)
1813                         break;
1814                 if (state->end >= cur_start && (state->state & bits) == bits) {
1815                         total_bytes += min(search_end, state->end) + 1 -
1816                                        max(cur_start, state->start);
1817                         if (total_bytes >= max_bytes)
1818                                 break;
1819                         if (!found) {
1820                                 *start = max(cur_start, state->start);
1821                                 found = 1;
1822                         }
1823                         last = state->end;
1824                 } else if (contig && found) {
1825                         break;
1826                 }
1827                 node = rb_next(node);
1828                 if (!node)
1829                         break;
1830         }
1831 out:
1832         spin_unlock(&tree->lock);
1833         return total_bytes;
1834 }
1835
1836 /*
1837  * set the private field for a given byte offset in the tree.  If there isn't
1838  * an extent_state there already, this does nothing.
1839  */
1840 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1841                 struct io_failure_record *failrec)
1842 {
1843         struct rb_node *node;
1844         struct extent_state *state;
1845         int ret = 0;
1846
1847         spin_lock(&tree->lock);
1848         /*
1849          * this search will find all the extents that end after
1850          * our range starts.
1851          */
1852         node = tree_search(tree, start);
1853         if (!node) {
1854                 ret = -ENOENT;
1855                 goto out;
1856         }
1857         state = rb_entry(node, struct extent_state, rb_node);
1858         if (state->start != start) {
1859                 ret = -ENOENT;
1860                 goto out;
1861         }
1862         state->failrec = failrec;
1863 out:
1864         spin_unlock(&tree->lock);
1865         return ret;
1866 }
1867
1868 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1869                 struct io_failure_record **failrec)
1870 {
1871         struct rb_node *node;
1872         struct extent_state *state;
1873         int ret = 0;
1874
1875         spin_lock(&tree->lock);
1876         /*
1877          * this search will find all the extents that end after
1878          * our range starts.
1879          */
1880         node = tree_search(tree, start);
1881         if (!node) {
1882                 ret = -ENOENT;
1883                 goto out;
1884         }
1885         state = rb_entry(node, struct extent_state, rb_node);
1886         if (state->start != start) {
1887                 ret = -ENOENT;
1888                 goto out;
1889         }
1890         *failrec = state->failrec;
1891 out:
1892         spin_unlock(&tree->lock);
1893         return ret;
1894 }
1895
1896 /*
1897  * searches a range in the state tree for a given mask.
1898  * If 'filled' == 1, this returns 1 only if every extent in the tree
1899  * has the bits set.  Otherwise, 1 is returned if any bit in the
1900  * range is found set.
1901  */
1902 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1903                    unsigned bits, int filled, struct extent_state *cached)
1904 {
1905         struct extent_state *state = NULL;
1906         struct rb_node *node;
1907         int bitset = 0;
1908
1909         spin_lock(&tree->lock);
1910         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1911             cached->end > start)
1912                 node = &cached->rb_node;
1913         else
1914                 node = tree_search(tree, start);
1915         while (node && start <= end) {
1916                 state = rb_entry(node, struct extent_state, rb_node);
1917
1918                 if (filled && state->start > start) {
1919                         bitset = 0;
1920                         break;
1921                 }
1922
1923                 if (state->start > end)
1924                         break;
1925
1926                 if (state->state & bits) {
1927                         bitset = 1;
1928                         if (!filled)
1929                                 break;
1930                 } else if (filled) {
1931                         bitset = 0;
1932                         break;
1933                 }
1934
1935                 if (state->end == (u64)-1)
1936                         break;
1937
1938                 start = state->end + 1;
1939                 if (start > end)
1940                         break;
1941                 node = rb_next(node);
1942                 if (!node) {
1943                         if (filled)
1944                                 bitset = 0;
1945                         break;
1946                 }
1947         }
1948         spin_unlock(&tree->lock);
1949         return bitset;
1950 }
1951
1952 /*
1953  * helper function to set a given page up to date if all the
1954  * extents in the tree for that page are up to date
1955  */
1956 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1957 {
1958         u64 start = page_offset(page);
1959         u64 end = start + PAGE_SIZE - 1;
1960         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1961                 SetPageUptodate(page);
1962 }
1963
1964 int free_io_failure(struct btrfs_inode *inode, struct io_failure_record *rec)
1965 {
1966         int ret;
1967         int err = 0;
1968         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
1969
1970         set_state_failrec(failure_tree, rec->start, NULL);
1971         ret = clear_extent_bits(failure_tree, rec->start,
1972                                 rec->start + rec->len - 1,
1973                                 EXTENT_LOCKED | EXTENT_DIRTY);
1974         if (ret)
1975                 err = ret;
1976
1977         ret = clear_extent_bits(&inode->io_tree, rec->start,
1978                                 rec->start + rec->len - 1,
1979                                 EXTENT_DAMAGED);
1980         if (ret && !err)
1981                 err = ret;
1982
1983         kfree(rec);
1984         return err;
1985 }
1986
1987 /*
1988  * this bypasses the standard btrfs submit functions deliberately, as
1989  * the standard behavior is to write all copies in a raid setup. here we only
1990  * want to write the one bad copy. so we do the mapping for ourselves and issue
1991  * submit_bio directly.
1992  * to avoid any synchronization issues, wait for the data after writing, which
1993  * actually prevents the read that triggered the error from finishing.
1994  * currently, there can be no more than two copies of every data bit. thus,
1995  * exactly one rewrite is required.
1996  */
1997 int repair_io_failure(struct btrfs_inode *inode, u64 start, u64 length,
1998                 u64 logical, struct page *page,
1999                 unsigned int pg_offset, int mirror_num)
2000 {
2001         struct btrfs_fs_info *fs_info = inode->root->fs_info;
2002         struct bio *bio;
2003         struct btrfs_device *dev;
2004         u64 map_length = 0;
2005         u64 sector;
2006         struct btrfs_bio *bbio = NULL;
2007         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2008         int ret;
2009
2010         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2011         BUG_ON(!mirror_num);
2012
2013         /* we can't repair anything in raid56 yet */
2014         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2015                 return 0;
2016
2017         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2018         if (!bio)
2019                 return -EIO;
2020         bio->bi_iter.bi_size = 0;
2021         map_length = length;
2022
2023         /*
2024          * Avoid races with device replace and make sure our bbio has devices
2025          * associated to its stripes that don't go away while we are doing the
2026          * read repair operation.
2027          */
2028         btrfs_bio_counter_inc_blocked(fs_info);
2029         ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2030                               &map_length, &bbio, mirror_num);
2031         if (ret) {
2032                 btrfs_bio_counter_dec(fs_info);
2033                 bio_put(bio);
2034                 return -EIO;
2035         }
2036         BUG_ON(mirror_num != bbio->mirror_num);
2037         sector = bbio->stripes[mirror_num-1].physical >> 9;
2038         bio->bi_iter.bi_sector = sector;
2039         dev = bbio->stripes[mirror_num-1].dev;
2040         btrfs_put_bbio(bbio);
2041         if (!dev || !dev->bdev || !dev->writeable) {
2042                 btrfs_bio_counter_dec(fs_info);
2043                 bio_put(bio);
2044                 return -EIO;
2045         }
2046         bio->bi_bdev = dev->bdev;
2047         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2048         bio_add_page(bio, page, length, pg_offset);
2049
2050         if (btrfsic_submit_bio_wait(bio)) {
2051                 /* try to remap that extent elsewhere? */
2052                 btrfs_bio_counter_dec(fs_info);
2053                 bio_put(bio);
2054                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2055                 return -EIO;
2056         }
2057
2058         btrfs_info_rl_in_rcu(fs_info,
2059                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2060                                   btrfs_ino(inode), start,
2061                                   rcu_str_deref(dev->name), sector);
2062         btrfs_bio_counter_dec(fs_info);
2063         bio_put(bio);
2064         return 0;
2065 }
2066
2067 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2068                          struct extent_buffer *eb, int mirror_num)
2069 {
2070         u64 start = eb->start;
2071         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2072         int ret = 0;
2073
2074         if (fs_info->sb->s_flags & MS_RDONLY)
2075                 return -EROFS;
2076
2077         for (i = 0; i < num_pages; i++) {
2078                 struct page *p = eb->pages[i];
2079
2080                 ret = repair_io_failure(BTRFS_I(fs_info->btree_inode), start,
2081                                         PAGE_SIZE, start, p,
2082                                         start - page_offset(p), mirror_num);
2083                 if (ret)
2084                         break;
2085                 start += PAGE_SIZE;
2086         }
2087
2088         return ret;
2089 }
2090
2091 /*
2092  * each time an IO finishes, we do a fast check in the IO failure tree
2093  * to see if we need to process or clean up an io_failure_record
2094  */
2095 int clean_io_failure(struct btrfs_inode *inode, u64 start, struct page *page,
2096                      unsigned int pg_offset)
2097 {
2098         u64 private;
2099         struct io_failure_record *failrec;
2100         struct btrfs_fs_info *fs_info = inode->root->fs_info;
2101         struct extent_state *state;
2102         int num_copies;
2103         int ret;
2104
2105         private = 0;
2106         ret = count_range_bits(&inode->io_failure_tree, &private,
2107                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2108         if (!ret)
2109                 return 0;
2110
2111         ret = get_state_failrec(&inode->io_failure_tree, start,
2112                         &failrec);
2113         if (ret)
2114                 return 0;
2115
2116         BUG_ON(!failrec->this_mirror);
2117
2118         if (failrec->in_validation) {
2119                 /* there was no real error, just free the record */
2120                 btrfs_debug(fs_info,
2121                         "clean_io_failure: freeing dummy error at %llu",
2122                         failrec->start);
2123                 goto out;
2124         }
2125         if (fs_info->sb->s_flags & MS_RDONLY)
2126                 goto out;
2127
2128         spin_lock(&inode->io_tree.lock);
2129         state = find_first_extent_bit_state(&inode->io_tree,
2130                                             failrec->start,
2131                                             EXTENT_LOCKED);
2132         spin_unlock(&inode->io_tree.lock);
2133
2134         if (state && state->start <= failrec->start &&
2135             state->end >= failrec->start + failrec->len - 1) {
2136                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2137                                               failrec->len);
2138                 if (num_copies > 1)  {
2139                         repair_io_failure(inode, start, failrec->len,
2140                                           failrec->logical, page,
2141                                           pg_offset, failrec->failed_mirror);
2142                 }
2143         }
2144
2145 out:
2146         free_io_failure(inode, failrec);
2147
2148         return 0;
2149 }
2150
2151 /*
2152  * Can be called when
2153  * - hold extent lock
2154  * - under ordered extent
2155  * - the inode is freeing
2156  */
2157 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2158 {
2159         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2160         struct io_failure_record *failrec;
2161         struct extent_state *state, *next;
2162
2163         if (RB_EMPTY_ROOT(&failure_tree->state))
2164                 return;
2165
2166         spin_lock(&failure_tree->lock);
2167         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2168         while (state) {
2169                 if (state->start > end)
2170                         break;
2171
2172                 ASSERT(state->end <= end);
2173
2174                 next = next_state(state);
2175
2176                 failrec = state->failrec;
2177                 free_extent_state(state);
2178                 kfree(failrec);
2179
2180                 state = next;
2181         }
2182         spin_unlock(&failure_tree->lock);
2183 }
2184
2185 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2186                 struct io_failure_record **failrec_ret)
2187 {
2188         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2189         struct io_failure_record *failrec;
2190         struct extent_map *em;
2191         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2192         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2193         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2194         int ret;
2195         u64 logical;
2196
2197         ret = get_state_failrec(failure_tree, start, &failrec);
2198         if (ret) {
2199                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2200                 if (!failrec)
2201                         return -ENOMEM;
2202
2203                 failrec->start = start;
2204                 failrec->len = end - start + 1;
2205                 failrec->this_mirror = 0;
2206                 failrec->bio_flags = 0;
2207                 failrec->in_validation = 0;
2208
2209                 read_lock(&em_tree->lock);
2210                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2211                 if (!em) {
2212                         read_unlock(&em_tree->lock);
2213                         kfree(failrec);
2214                         return -EIO;
2215                 }
2216
2217                 if (em->start > start || em->start + em->len <= start) {
2218                         free_extent_map(em);
2219                         em = NULL;
2220                 }
2221                 read_unlock(&em_tree->lock);
2222                 if (!em) {
2223                         kfree(failrec);
2224                         return -EIO;
2225                 }
2226
2227                 logical = start - em->start;
2228                 logical = em->block_start + logical;
2229                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2230                         logical = em->block_start;
2231                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2232                         extent_set_compress_type(&failrec->bio_flags,
2233                                                  em->compress_type);
2234                 }
2235
2236                 btrfs_debug(fs_info,
2237                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2238                         logical, start, failrec->len);
2239
2240                 failrec->logical = logical;
2241                 free_extent_map(em);
2242
2243                 /* set the bits in the private failure tree */
2244                 ret = set_extent_bits(failure_tree, start, end,
2245                                         EXTENT_LOCKED | EXTENT_DIRTY);
2246                 if (ret >= 0)
2247                         ret = set_state_failrec(failure_tree, start, failrec);
2248                 /* set the bits in the inode's tree */
2249                 if (ret >= 0)
2250                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2251                 if (ret < 0) {
2252                         kfree(failrec);
2253                         return ret;
2254                 }
2255         } else {
2256                 btrfs_debug(fs_info,
2257                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2258                         failrec->logical, failrec->start, failrec->len,
2259                         failrec->in_validation);
2260                 /*
2261                  * when data can be on disk more than twice, add to failrec here
2262                  * (e.g. with a list for failed_mirror) to make
2263                  * clean_io_failure() clean all those errors at once.
2264                  */
2265         }
2266
2267         *failrec_ret = failrec;
2268
2269         return 0;
2270 }
2271
2272 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2273                            struct io_failure_record *failrec, int failed_mirror)
2274 {
2275         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2276         int num_copies;
2277
2278         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2279         if (num_copies == 1) {
2280                 /*
2281                  * we only have a single copy of the data, so don't bother with
2282                  * all the retry and error correction code that follows. no
2283                  * matter what the error is, it is very likely to persist.
2284                  */
2285                 btrfs_debug(fs_info,
2286                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2287                         num_copies, failrec->this_mirror, failed_mirror);
2288                 return 0;
2289         }
2290
2291         /*
2292          * there are two premises:
2293          *      a) deliver good data to the caller
2294          *      b) correct the bad sectors on disk
2295          */
2296         if (failed_bio->bi_vcnt > 1) {
2297                 /*
2298                  * to fulfill b), we need to know the exact failing sectors, as
2299                  * we don't want to rewrite any more than the failed ones. thus,
2300                  * we need separate read requests for the failed bio
2301                  *
2302                  * if the following BUG_ON triggers, our validation request got
2303                  * merged. we need separate requests for our algorithm to work.
2304                  */
2305                 BUG_ON(failrec->in_validation);
2306                 failrec->in_validation = 1;
2307                 failrec->this_mirror = failed_mirror;
2308         } else {
2309                 /*
2310                  * we're ready to fulfill a) and b) alongside. get a good copy
2311                  * of the failed sector and if we succeed, we have setup
2312                  * everything for repair_io_failure to do the rest for us.
2313                  */
2314                 if (failrec->in_validation) {
2315                         BUG_ON(failrec->this_mirror != failed_mirror);
2316                         failrec->in_validation = 0;
2317                         failrec->this_mirror = 0;
2318                 }
2319                 failrec->failed_mirror = failed_mirror;
2320                 failrec->this_mirror++;
2321                 if (failrec->this_mirror == failed_mirror)
2322                         failrec->this_mirror++;
2323         }
2324
2325         if (failrec->this_mirror > num_copies) {
2326                 btrfs_debug(fs_info,
2327                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2328                         num_copies, failrec->this_mirror, failed_mirror);
2329                 return 0;
2330         }
2331
2332         return 1;
2333 }
2334
2335
2336 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2337                                     struct io_failure_record *failrec,
2338                                     struct page *page, int pg_offset, int icsum,
2339                                     bio_end_io_t *endio_func, void *data)
2340 {
2341         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2342         struct bio *bio;
2343         struct btrfs_io_bio *btrfs_failed_bio;
2344         struct btrfs_io_bio *btrfs_bio;
2345
2346         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2347         if (!bio)
2348                 return NULL;
2349
2350         bio->bi_end_io = endio_func;
2351         bio->bi_iter.bi_sector = failrec->logical >> 9;
2352         bio->bi_bdev = fs_info->fs_devices->latest_bdev;
2353         bio->bi_iter.bi_size = 0;
2354         bio->bi_private = data;
2355
2356         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2357         if (btrfs_failed_bio->csum) {
2358                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2359
2360                 btrfs_bio = btrfs_io_bio(bio);
2361                 btrfs_bio->csum = btrfs_bio->csum_inline;
2362                 icsum *= csum_size;
2363                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2364                        csum_size);
2365         }
2366
2367         bio_add_page(bio, page, failrec->len, pg_offset);
2368
2369         return bio;
2370 }
2371
2372 /*
2373  * this is a generic handler for readpage errors (default
2374  * readpage_io_failed_hook). if other copies exist, read those and write back
2375  * good data to the failed position. does not investigate in remapping the
2376  * failed extent elsewhere, hoping the device will be smart enough to do this as
2377  * needed
2378  */
2379
2380 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2381                               struct page *page, u64 start, u64 end,
2382                               int failed_mirror)
2383 {
2384         struct io_failure_record *failrec;
2385         struct inode *inode = page->mapping->host;
2386         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2387         struct bio *bio;
2388         int read_mode = 0;
2389         int ret;
2390
2391         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2392
2393         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2394         if (ret)
2395                 return ret;
2396
2397         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2398         if (!ret) {
2399                 free_io_failure(BTRFS_I(inode), failrec);
2400                 return -EIO;
2401         }
2402
2403         if (failed_bio->bi_vcnt > 1)
2404                 read_mode |= REQ_FAILFAST_DEV;
2405
2406         phy_offset >>= inode->i_sb->s_blocksize_bits;
2407         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2408                                       start - page_offset(page),
2409                                       (int)phy_offset, failed_bio->bi_end_io,
2410                                       NULL);
2411         if (!bio) {
2412                 free_io_failure(BTRFS_I(inode), failrec);
2413                 return -EIO;
2414         }
2415         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2416
2417         btrfs_debug(btrfs_sb(inode->i_sb),
2418                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2419                 read_mode, failrec->this_mirror, failrec->in_validation);
2420
2421         ret = tree->ops->submit_bio_hook(inode, bio, failrec->this_mirror,
2422                                          failrec->bio_flags, 0);
2423         if (ret) {
2424                 free_io_failure(BTRFS_I(inode), failrec);
2425                 bio_put(bio);
2426         }
2427
2428         return ret;
2429 }
2430
2431 /* lots and lots of room for performance fixes in the end_bio funcs */
2432
2433 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2434 {
2435         int uptodate = (err == 0);
2436         struct extent_io_tree *tree;
2437         int ret = 0;
2438
2439         tree = &BTRFS_I(page->mapping->host)->io_tree;
2440
2441         if (tree->ops && tree->ops->writepage_end_io_hook)
2442                 tree->ops->writepage_end_io_hook(page, start, end, NULL,
2443                                 uptodate);
2444
2445         if (!uptodate) {
2446                 ClearPageUptodate(page);
2447                 SetPageError(page);
2448                 ret = ret < 0 ? ret : -EIO;
2449                 mapping_set_error(page->mapping, ret);
2450         }
2451 }
2452
2453 /*
2454  * after a writepage IO is done, we need to:
2455  * clear the uptodate bits on error
2456  * clear the writeback bits in the extent tree for this IO
2457  * end_page_writeback if the page has no more pending IO
2458  *
2459  * Scheduling is not allowed, so the extent state tree is expected
2460  * to have one and only one object corresponding to this IO.
2461  */
2462 static void end_bio_extent_writepage(struct bio *bio)
2463 {
2464         struct bio_vec *bvec;
2465         u64 start;
2466         u64 end;
2467         int i;
2468
2469         bio_for_each_segment_all(bvec, bio, i) {
2470                 struct page *page = bvec->bv_page;
2471                 struct inode *inode = page->mapping->host;
2472                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2473
2474                 /* We always issue full-page reads, but if some block
2475                  * in a page fails to read, blk_update_request() will
2476                  * advance bv_offset and adjust bv_len to compensate.
2477                  * Print a warning for nonzero offsets, and an error
2478                  * if they don't add up to a full page.  */
2479                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2480                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2481                                 btrfs_err(fs_info,
2482                                    "partial page write in btrfs with offset %u and length %u",
2483                                         bvec->bv_offset, bvec->bv_len);
2484                         else
2485                                 btrfs_info(fs_info,
2486                                    "incomplete page write in btrfs with offset %u and length %u",
2487                                         bvec->bv_offset, bvec->bv_len);
2488                 }
2489
2490                 start = page_offset(page);
2491                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2492
2493                 end_extent_writepage(page, bio->bi_error, start, end);
2494                 end_page_writeback(page);
2495         }
2496
2497         bio_put(bio);
2498 }
2499
2500 static void
2501 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2502                               int uptodate)
2503 {
2504         struct extent_state *cached = NULL;
2505         u64 end = start + len - 1;
2506
2507         if (uptodate && tree->track_uptodate)
2508                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2509         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2510 }
2511
2512 /*
2513  * after a readpage IO is done, we need to:
2514  * clear the uptodate bits on error
2515  * set the uptodate bits if things worked
2516  * set the page up to date if all extents in the tree are uptodate
2517  * clear the lock bit in the extent tree
2518  * unlock the page if there are no other extents locked for it
2519  *
2520  * Scheduling is not allowed, so the extent state tree is expected
2521  * to have one and only one object corresponding to this IO.
2522  */
2523 static void end_bio_extent_readpage(struct bio *bio)
2524 {
2525         struct bio_vec *bvec;
2526         int uptodate = !bio->bi_error;
2527         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2528         struct extent_io_tree *tree;
2529         u64 offset = 0;
2530         u64 start;
2531         u64 end;
2532         u64 len;
2533         u64 extent_start = 0;
2534         u64 extent_len = 0;
2535         int mirror;
2536         int ret;
2537         int i;
2538
2539         bio_for_each_segment_all(bvec, bio, i) {
2540                 struct page *page = bvec->bv_page;
2541                 struct inode *inode = page->mapping->host;
2542                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2543
2544                 btrfs_debug(fs_info,
2545                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2546                         (u64)bio->bi_iter.bi_sector, bio->bi_error,
2547                         io_bio->mirror_num);
2548                 tree = &BTRFS_I(inode)->io_tree;
2549
2550                 /* We always issue full-page reads, but if some block
2551                  * in a page fails to read, blk_update_request() will
2552                  * advance bv_offset and adjust bv_len to compensate.
2553                  * Print a warning for nonzero offsets, and an error
2554                  * if they don't add up to a full page.  */
2555                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2556                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2557                                 btrfs_err(fs_info,
2558                                         "partial page read in btrfs with offset %u and length %u",
2559                                         bvec->bv_offset, bvec->bv_len);
2560                         else
2561                                 btrfs_info(fs_info,
2562                                         "incomplete page read in btrfs with offset %u and length %u",
2563                                         bvec->bv_offset, bvec->bv_len);
2564                 }
2565
2566                 start = page_offset(page);
2567                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2568                 len = bvec->bv_len;
2569
2570                 mirror = io_bio->mirror_num;
2571                 if (likely(uptodate && tree->ops)) {
2572                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2573                                                               page, start, end,
2574                                                               mirror);
2575                         if (ret)
2576                                 uptodate = 0;
2577                         else
2578                                 clean_io_failure(BTRFS_I(inode), start,
2579                                                 page, 0);
2580                 }
2581
2582                 if (likely(uptodate))
2583                         goto readpage_ok;
2584
2585                 if (tree->ops) {
2586                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2587                         if (ret == -EAGAIN) {
2588                                 /*
2589                                  * Data inode's readpage_io_failed_hook() always
2590                                  * returns -EAGAIN.
2591                                  *
2592                                  * The generic bio_readpage_error handles errors
2593                                  * the following way: If possible, new read
2594                                  * requests are created and submitted and will
2595                                  * end up in end_bio_extent_readpage as well (if
2596                                  * we're lucky, not in the !uptodate case). In
2597                                  * that case it returns 0 and we just go on with
2598                                  * the next page in our bio. If it can't handle
2599                                  * the error it will return -EIO and we remain
2600                                  * responsible for that page.
2601                                  */
2602                                 ret = bio_readpage_error(bio, offset, page,
2603                                                          start, end, mirror);
2604                                 if (ret == 0) {
2605                                         uptodate = !bio->bi_error;
2606                                         offset += len;
2607                                         continue;
2608                                 }
2609                         }
2610
2611                         /*
2612                          * metadata's readpage_io_failed_hook() always returns
2613                          * -EIO and fixes nothing.  -EIO is also returned if
2614                          * data inode error could not be fixed.
2615                          */
2616                         ASSERT(ret == -EIO);
2617                 }
2618 readpage_ok:
2619                 if (likely(uptodate)) {
2620                         loff_t i_size = i_size_read(inode);
2621                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2622                         unsigned off;
2623
2624                         /* Zero out the end if this page straddles i_size */
2625                         off = i_size & (PAGE_SIZE-1);
2626                         if (page->index == end_index && off)
2627                                 zero_user_segment(page, off, PAGE_SIZE);
2628                         SetPageUptodate(page);
2629                 } else {
2630                         ClearPageUptodate(page);
2631                         SetPageError(page);
2632                 }
2633                 unlock_page(page);
2634                 offset += len;
2635
2636                 if (unlikely(!uptodate)) {
2637                         if (extent_len) {
2638                                 endio_readpage_release_extent(tree,
2639                                                               extent_start,
2640                                                               extent_len, 1);
2641                                 extent_start = 0;
2642                                 extent_len = 0;
2643                         }
2644                         endio_readpage_release_extent(tree, start,
2645                                                       end - start + 1, 0);
2646                 } else if (!extent_len) {
2647                         extent_start = start;
2648                         extent_len = end + 1 - start;
2649                 } else if (extent_start + extent_len == start) {
2650                         extent_len += end + 1 - start;
2651                 } else {
2652                         endio_readpage_release_extent(tree, extent_start,
2653                                                       extent_len, uptodate);
2654                         extent_start = start;
2655                         extent_len = end + 1 - start;
2656                 }
2657         }
2658
2659         if (extent_len)
2660                 endio_readpage_release_extent(tree, extent_start, extent_len,
2661                                               uptodate);
2662         if (io_bio->end_io)
2663                 io_bio->end_io(io_bio, bio->bi_error);
2664         bio_put(bio);
2665 }
2666
2667 /*
2668  * this allocates from the btrfs_bioset.  We're returning a bio right now
2669  * but you can call btrfs_io_bio for the appropriate container_of magic
2670  */
2671 struct bio *
2672 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2673                 gfp_t gfp_flags)
2674 {
2675         struct btrfs_io_bio *btrfs_bio;
2676         struct bio *bio;
2677
2678         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2679
2680         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2681                 while (!bio && (nr_vecs /= 2)) {
2682                         bio = bio_alloc_bioset(gfp_flags,
2683                                                nr_vecs, btrfs_bioset);
2684                 }
2685         }
2686
2687         if (bio) {
2688                 bio->bi_bdev = bdev;
2689                 bio->bi_iter.bi_sector = first_sector;
2690                 btrfs_bio = btrfs_io_bio(bio);
2691                 btrfs_bio->csum = NULL;
2692                 btrfs_bio->csum_allocated = NULL;
2693                 btrfs_bio->end_io = NULL;
2694         }
2695         return bio;
2696 }
2697
2698 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2699 {
2700         struct btrfs_io_bio *btrfs_bio;
2701         struct bio *new;
2702
2703         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2704         if (new) {
2705                 btrfs_bio = btrfs_io_bio(new);
2706                 btrfs_bio->csum = NULL;
2707                 btrfs_bio->csum_allocated = NULL;
2708                 btrfs_bio->end_io = NULL;
2709         }
2710         return new;
2711 }
2712
2713 /* this also allocates from the btrfs_bioset */
2714 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2715 {
2716         struct btrfs_io_bio *btrfs_bio;
2717         struct bio *bio;
2718
2719         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2720         if (bio) {
2721                 btrfs_bio = btrfs_io_bio(bio);
2722                 btrfs_bio->csum = NULL;
2723                 btrfs_bio->csum_allocated = NULL;
2724                 btrfs_bio->end_io = NULL;
2725         }
2726         return bio;
2727 }
2728
2729
2730 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2731                                        unsigned long bio_flags)
2732 {
2733         int ret = 0;
2734         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2735         struct page *page = bvec->bv_page;
2736         struct extent_io_tree *tree = bio->bi_private;
2737         u64 start;
2738
2739         start = page_offset(page) + bvec->bv_offset;
2740
2741         bio->bi_private = NULL;
2742         bio_get(bio);
2743
2744         if (tree->ops)
2745                 ret = tree->ops->submit_bio_hook(page->mapping->host, bio,
2746                                            mirror_num, bio_flags, start);
2747         else
2748                 btrfsic_submit_bio(bio);
2749
2750         bio_put(bio);
2751         return ret;
2752 }
2753
2754 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2755                      unsigned long offset, size_t size, struct bio *bio,
2756                      unsigned long bio_flags)
2757 {
2758         int ret = 0;
2759         if (tree->ops)
2760                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2761                                                 bio_flags);
2762         return ret;
2763
2764 }
2765
2766 static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree,
2767                               struct writeback_control *wbc,
2768                               struct page *page, sector_t sector,
2769                               size_t size, unsigned long offset,
2770                               struct block_device *bdev,
2771                               struct bio **bio_ret,
2772                               bio_end_io_t end_io_func,
2773                               int mirror_num,
2774                               unsigned long prev_bio_flags,
2775                               unsigned long bio_flags,
2776                               bool force_bio_submit)
2777 {
2778         int ret = 0;
2779         struct bio *bio;
2780         int contig = 0;
2781         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2782         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2783
2784         if (bio_ret && *bio_ret) {
2785                 bio = *bio_ret;
2786                 if (old_compressed)
2787                         contig = bio->bi_iter.bi_sector == sector;
2788                 else
2789                         contig = bio_end_sector(bio) == sector;
2790
2791                 if (prev_bio_flags != bio_flags || !contig ||
2792                     force_bio_submit ||
2793                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2794                     bio_add_page(bio, page, page_size, offset) < page_size) {
2795                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2796                         if (ret < 0) {
2797                                 *bio_ret = NULL;
2798                                 return ret;
2799                         }
2800                         bio = NULL;
2801                 } else {
2802                         if (wbc)
2803                                 wbc_account_io(wbc, page, page_size);
2804                         return 0;
2805                 }
2806         }
2807
2808         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2809                         GFP_NOFS | __GFP_HIGH);
2810         if (!bio)
2811                 return -ENOMEM;
2812
2813         bio_add_page(bio, page, page_size, offset);
2814         bio->bi_end_io = end_io_func;
2815         bio->bi_private = tree;
2816         bio_set_op_attrs(bio, op, op_flags);
2817         if (wbc) {
2818                 wbc_init_bio(wbc, bio);
2819                 wbc_account_io(wbc, page, page_size);
2820         }
2821
2822         if (bio_ret)
2823                 *bio_ret = bio;
2824         else
2825                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2826
2827         return ret;
2828 }
2829
2830 static void attach_extent_buffer_page(struct extent_buffer *eb,
2831                                       struct page *page)
2832 {
2833         if (!PagePrivate(page)) {
2834                 SetPagePrivate(page);
2835                 get_page(page);
2836                 set_page_private(page, (unsigned long)eb);
2837         } else {
2838                 WARN_ON(page->private != (unsigned long)eb);
2839         }
2840 }
2841
2842 void set_page_extent_mapped(struct page *page)
2843 {
2844         if (!PagePrivate(page)) {
2845                 SetPagePrivate(page);
2846                 get_page(page);
2847                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2848         }
2849 }
2850
2851 static struct extent_map *
2852 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2853                  u64 start, u64 len, get_extent_t *get_extent,
2854                  struct extent_map **em_cached)
2855 {
2856         struct extent_map *em;
2857
2858         if (em_cached && *em_cached) {
2859                 em = *em_cached;
2860                 if (extent_map_in_tree(em) && start >= em->start &&
2861                     start < extent_map_end(em)) {
2862                         atomic_inc(&em->refs);
2863                         return em;
2864                 }
2865
2866                 free_extent_map(em);
2867                 *em_cached = NULL;
2868         }
2869
2870         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2871         if (em_cached && !IS_ERR_OR_NULL(em)) {
2872                 BUG_ON(*em_cached);
2873                 atomic_inc(&em->refs);
2874                 *em_cached = em;
2875         }
2876         return em;
2877 }
2878 /*
2879  * basic readpage implementation.  Locked extent state structs are inserted
2880  * into the tree that are removed when the IO is done (by the end_io
2881  * handlers)
2882  * XXX JDM: This needs looking at to ensure proper page locking
2883  * return 0 on success, otherwise return error
2884  */
2885 static int __do_readpage(struct extent_io_tree *tree,
2886                          struct page *page,
2887                          get_extent_t *get_extent,
2888                          struct extent_map **em_cached,
2889                          struct bio **bio, int mirror_num,
2890                          unsigned long *bio_flags, int read_flags,
2891                          u64 *prev_em_start)
2892 {
2893         struct inode *inode = page->mapping->host;
2894         u64 start = page_offset(page);
2895         u64 page_end = start + PAGE_SIZE - 1;
2896         u64 end;
2897         u64 cur = start;
2898         u64 extent_offset;
2899         u64 last_byte = i_size_read(inode);
2900         u64 block_start;
2901         u64 cur_end;
2902         sector_t sector;
2903         struct extent_map *em;
2904         struct block_device *bdev;
2905         int ret = 0;
2906         int nr = 0;
2907         size_t pg_offset = 0;
2908         size_t iosize;
2909         size_t disk_io_size;
2910         size_t blocksize = inode->i_sb->s_blocksize;
2911         unsigned long this_bio_flag = 0;
2912
2913         set_page_extent_mapped(page);
2914
2915         end = page_end;
2916         if (!PageUptodate(page)) {
2917                 if (cleancache_get_page(page) == 0) {
2918                         BUG_ON(blocksize != PAGE_SIZE);
2919                         unlock_extent(tree, start, end);
2920                         goto out;
2921                 }
2922         }
2923
2924         if (page->index == last_byte >> PAGE_SHIFT) {
2925                 char *userpage;
2926                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2927
2928                 if (zero_offset) {
2929                         iosize = PAGE_SIZE - zero_offset;
2930                         userpage = kmap_atomic(page);
2931                         memset(userpage + zero_offset, 0, iosize);
2932                         flush_dcache_page(page);
2933                         kunmap_atomic(userpage);
2934                 }
2935         }
2936         while (cur <= end) {
2937                 bool force_bio_submit = false;
2938
2939                 if (cur >= last_byte) {
2940                         char *userpage;
2941                         struct extent_state *cached = NULL;
2942
2943                         iosize = PAGE_SIZE - pg_offset;
2944                         userpage = kmap_atomic(page);
2945                         memset(userpage + pg_offset, 0, iosize);
2946                         flush_dcache_page(page);
2947                         kunmap_atomic(userpage);
2948                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2949                                             &cached, GFP_NOFS);
2950                         unlock_extent_cached(tree, cur,
2951                                              cur + iosize - 1,
2952                                              &cached, GFP_NOFS);
2953                         break;
2954                 }
2955                 em = __get_extent_map(inode, page, pg_offset, cur,
2956                                       end - cur + 1, get_extent, em_cached);
2957                 if (IS_ERR_OR_NULL(em)) {
2958                         SetPageError(page);
2959                         unlock_extent(tree, cur, end);
2960                         break;
2961                 }
2962                 extent_offset = cur - em->start;
2963                 BUG_ON(extent_map_end(em) <= cur);
2964                 BUG_ON(end < cur);
2965
2966                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2967                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2968                         extent_set_compress_type(&this_bio_flag,
2969                                                  em->compress_type);
2970                 }
2971
2972                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2973                 cur_end = min(extent_map_end(em) - 1, end);
2974                 iosize = ALIGN(iosize, blocksize);
2975                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2976                         disk_io_size = em->block_len;
2977                         sector = em->block_start >> 9;
2978                 } else {
2979                         sector = (em->block_start + extent_offset) >> 9;
2980                         disk_io_size = iosize;
2981                 }
2982                 bdev = em->bdev;
2983                 block_start = em->block_start;
2984                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2985                         block_start = EXTENT_MAP_HOLE;
2986
2987                 /*
2988                  * If we have a file range that points to a compressed extent
2989                  * and it's followed by a consecutive file range that points to
2990                  * to the same compressed extent (possibly with a different
2991                  * offset and/or length, so it either points to the whole extent
2992                  * or only part of it), we must make sure we do not submit a
2993                  * single bio to populate the pages for the 2 ranges because
2994                  * this makes the compressed extent read zero out the pages
2995                  * belonging to the 2nd range. Imagine the following scenario:
2996                  *
2997                  *  File layout
2998                  *  [0 - 8K]                     [8K - 24K]
2999                  *    |                               |
3000                  *    |                               |
3001                  * points to extent X,         points to extent X,
3002                  * offset 4K, length of 8K     offset 0, length 16K
3003                  *
3004                  * [extent X, compressed length = 4K uncompressed length = 16K]
3005                  *
3006                  * If the bio to read the compressed extent covers both ranges,
3007                  * it will decompress extent X into the pages belonging to the
3008                  * first range and then it will stop, zeroing out the remaining
3009                  * pages that belong to the other range that points to extent X.
3010                  * So here we make sure we submit 2 bios, one for the first
3011                  * range and another one for the third range. Both will target
3012                  * the same physical extent from disk, but we can't currently
3013                  * make the compressed bio endio callback populate the pages
3014                  * for both ranges because each compressed bio is tightly
3015                  * coupled with a single extent map, and each range can have
3016                  * an extent map with a different offset value relative to the
3017                  * uncompressed data of our extent and different lengths. This
3018                  * is a corner case so we prioritize correctness over
3019                  * non-optimal behavior (submitting 2 bios for the same extent).
3020                  */
3021                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3022                     prev_em_start && *prev_em_start != (u64)-1 &&
3023                     *prev_em_start != em->orig_start)
3024                         force_bio_submit = true;
3025
3026                 if (prev_em_start)
3027                         *prev_em_start = em->orig_start;
3028
3029                 free_extent_map(em);
3030                 em = NULL;
3031
3032                 /* we've found a hole, just zero and go on */
3033                 if (block_start == EXTENT_MAP_HOLE) {
3034                         char *userpage;
3035                         struct extent_state *cached = NULL;
3036
3037                         userpage = kmap_atomic(page);
3038                         memset(userpage + pg_offset, 0, iosize);
3039                         flush_dcache_page(page);
3040                         kunmap_atomic(userpage);
3041
3042                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3043                                             &cached, GFP_NOFS);
3044                         unlock_extent_cached(tree, cur,
3045                                              cur + iosize - 1,
3046                                              &cached, GFP_NOFS);
3047                         cur = cur + iosize;
3048                         pg_offset += iosize;
3049                         continue;
3050                 }
3051                 /* the get_extent function already copied into the page */
3052                 if (test_range_bit(tree, cur, cur_end,
3053                                    EXTENT_UPTODATE, 1, NULL)) {
3054                         check_page_uptodate(tree, page);
3055                         unlock_extent(tree, cur, cur + iosize - 1);
3056                         cur = cur + iosize;
3057                         pg_offset += iosize;
3058                         continue;
3059                 }
3060                 /* we have an inline extent but it didn't get marked up
3061                  * to date.  Error out
3062                  */
3063                 if (block_start == EXTENT_MAP_INLINE) {
3064                         SetPageError(page);
3065                         unlock_extent(tree, cur, cur + iosize - 1);
3066                         cur = cur + iosize;
3067                         pg_offset += iosize;
3068                         continue;
3069                 }
3070
3071                 ret = submit_extent_page(REQ_OP_READ, read_flags, tree, NULL,
3072                                          page, sector, disk_io_size, pg_offset,
3073                                          bdev, bio,
3074                                          end_bio_extent_readpage, mirror_num,
3075                                          *bio_flags,
3076                                          this_bio_flag,
3077                                          force_bio_submit);
3078                 if (!ret) {
3079                         nr++;
3080                         *bio_flags = this_bio_flag;
3081                 } else {
3082                         SetPageError(page);
3083                         unlock_extent(tree, cur, cur + iosize - 1);
3084                         goto out;
3085                 }
3086                 cur = cur + iosize;
3087                 pg_offset += iosize;
3088         }
3089 out:
3090         if (!nr) {
3091                 if (!PageError(page))
3092                         SetPageUptodate(page);
3093                 unlock_page(page);
3094         }
3095         return ret;
3096 }
3097
3098 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3099                                              struct page *pages[], int nr_pages,
3100                                              u64 start, u64 end,
3101                                              get_extent_t *get_extent,
3102                                              struct extent_map **em_cached,
3103                                              struct bio **bio, int mirror_num,
3104                                              unsigned long *bio_flags,
3105                                              u64 *prev_em_start)
3106 {
3107         struct inode *inode;
3108         struct btrfs_ordered_extent *ordered;
3109         int index;
3110
3111         inode = pages[0]->mapping->host;
3112         while (1) {
3113                 lock_extent(tree, start, end);
3114                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3115                                                      end - start + 1);
3116                 if (!ordered)
3117                         break;
3118                 unlock_extent(tree, start, end);
3119                 btrfs_start_ordered_extent(inode, ordered, 1);
3120                 btrfs_put_ordered_extent(ordered);
3121         }
3122
3123         for (index = 0; index < nr_pages; index++) {
3124                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3125                               mirror_num, bio_flags, 0, prev_em_start);
3126                 put_page(pages[index]);
3127         }
3128 }
3129
3130 static void __extent_readpages(struct extent_io_tree *tree,
3131                                struct page *pages[],
3132                                int nr_pages, get_extent_t *get_extent,
3133                                struct extent_map **em_cached,
3134                                struct bio **bio, int mirror_num,
3135                                unsigned long *bio_flags,
3136                                u64 *prev_em_start)
3137 {
3138         u64 start = 0;
3139         u64 end = 0;
3140         u64 page_start;
3141         int index;
3142         int first_index = 0;
3143
3144         for (index = 0; index < nr_pages; index++) {
3145                 page_start = page_offset(pages[index]);
3146                 if (!end) {
3147                         start = page_start;
3148                         end = start + PAGE_SIZE - 1;
3149                         first_index = index;
3150                 } else if (end + 1 == page_start) {
3151                         end += PAGE_SIZE;
3152                 } else {
3153                         __do_contiguous_readpages(tree, &pages[first_index],
3154                                                   index - first_index, start,
3155                                                   end, get_extent, em_cached,
3156                                                   bio, mirror_num, bio_flags,
3157                                                   prev_em_start);
3158                         start = page_start;
3159                         end = start + PAGE_SIZE - 1;
3160                         first_index = index;
3161                 }
3162         }
3163
3164         if (end)
3165                 __do_contiguous_readpages(tree, &pages[first_index],
3166                                           index - first_index, start,
3167                                           end, get_extent, em_cached, bio,
3168                                           mirror_num, bio_flags,
3169                                           prev_em_start);
3170 }
3171
3172 static int __extent_read_full_page(struct extent_io_tree *tree,
3173                                    struct page *page,
3174                                    get_extent_t *get_extent,
3175                                    struct bio **bio, int mirror_num,
3176                                    unsigned long *bio_flags, int read_flags)
3177 {
3178         struct inode *inode = page->mapping->host;
3179         struct btrfs_ordered_extent *ordered;
3180         u64 start = page_offset(page);
3181         u64 end = start + PAGE_SIZE - 1;
3182         int ret;
3183
3184         while (1) {
3185                 lock_extent(tree, start, end);
3186                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3187                                                 PAGE_SIZE);
3188                 if (!ordered)
3189                         break;
3190                 unlock_extent(tree, start, end);
3191                 btrfs_start_ordered_extent(inode, ordered, 1);
3192                 btrfs_put_ordered_extent(ordered);
3193         }
3194
3195         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3196                             bio_flags, read_flags, NULL);
3197         return ret;
3198 }
3199
3200 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3201                             get_extent_t *get_extent, int mirror_num)
3202 {
3203         struct bio *bio = NULL;
3204         unsigned long bio_flags = 0;
3205         int ret;
3206
3207         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3208                                       &bio_flags, 0);
3209         if (bio)
3210                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3211         return ret;
3212 }
3213
3214 static void update_nr_written(struct writeback_control *wbc,
3215                               unsigned long nr_written)
3216 {
3217         wbc->nr_to_write -= nr_written;
3218 }
3219
3220 /*
3221  * helper for __extent_writepage, doing all of the delayed allocation setup.
3222  *
3223  * This returns 1 if our fill_delalloc function did all the work required
3224  * to write the page (copy into inline extent).  In this case the IO has
3225  * been started and the page is already unlocked.
3226  *
3227  * This returns 0 if all went well (page still locked)
3228  * This returns < 0 if there were errors (page still locked)
3229  */
3230 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3231                               struct page *page, struct writeback_control *wbc,
3232                               struct extent_page_data *epd,
3233                               u64 delalloc_start,
3234                               unsigned long *nr_written)
3235 {
3236         struct extent_io_tree *tree = epd->tree;
3237         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3238         u64 nr_delalloc;
3239         u64 delalloc_to_write = 0;
3240         u64 delalloc_end = 0;
3241         int ret;
3242         int page_started = 0;
3243
3244         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3245                 return 0;
3246
3247         while (delalloc_end < page_end) {
3248                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3249                                                page,
3250                                                &delalloc_start,
3251                                                &delalloc_end,
3252                                                BTRFS_MAX_EXTENT_SIZE);
3253                 if (nr_delalloc == 0) {
3254                         delalloc_start = delalloc_end + 1;
3255                         continue;
3256                 }
3257                 ret = tree->ops->fill_delalloc(inode, page,
3258                                                delalloc_start,
3259                                                delalloc_end,
3260                                                &page_started,
3261                                                nr_written);
3262                 /* File system has been set read-only */
3263                 if (ret) {
3264                         SetPageError(page);
3265                         /* fill_delalloc should be return < 0 for error
3266                          * but just in case, we use > 0 here meaning the
3267                          * IO is started, so we don't want to return > 0
3268                          * unless things are going well.
3269                          */
3270                         ret = ret < 0 ? ret : -EIO;
3271                         goto done;
3272                 }
3273                 /*
3274                  * delalloc_end is already one less than the total length, so
3275                  * we don't subtract one from PAGE_SIZE
3276                  */
3277                 delalloc_to_write += (delalloc_end - delalloc_start +
3278                                       PAGE_SIZE) >> PAGE_SHIFT;
3279                 delalloc_start = delalloc_end + 1;
3280         }
3281         if (wbc->nr_to_write < delalloc_to_write) {
3282                 int thresh = 8192;
3283
3284                 if (delalloc_to_write < thresh * 2)
3285                         thresh = delalloc_to_write;
3286                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3287                                          thresh);
3288         }
3289
3290         /* did the fill delalloc function already unlock and start
3291          * the IO?
3292          */
3293         if (page_started) {
3294                 /*
3295                  * we've unlocked the page, so we can't update
3296                  * the mapping's writeback index, just update
3297                  * nr_to_write.
3298                  */
3299                 wbc->nr_to_write -= *nr_written;
3300                 return 1;
3301         }
3302
3303         ret = 0;
3304
3305 done:
3306         return ret;
3307 }
3308
3309 /*
3310  * helper for __extent_writepage.  This calls the writepage start hooks,
3311  * and does the loop to map the page into extents and bios.
3312  *
3313  * We return 1 if the IO is started and the page is unlocked,
3314  * 0 if all went well (page still locked)
3315  * < 0 if there were errors (page still locked)
3316  */
3317 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3318                                  struct page *page,
3319                                  struct writeback_control *wbc,
3320                                  struct extent_page_data *epd,
3321                                  loff_t i_size,
3322                                  unsigned long nr_written,
3323                                  int write_flags, int *nr_ret)
3324 {
3325         struct extent_io_tree *tree = epd->tree;
3326         u64 start = page_offset(page);
3327         u64 page_end = start + PAGE_SIZE - 1;
3328         u64 end;
3329         u64 cur = start;
3330         u64 extent_offset;
3331         u64 block_start;
3332         u64 iosize;
3333         sector_t sector;
3334         struct extent_map *em;
3335         struct block_device *bdev;
3336         size_t pg_offset = 0;
3337         size_t blocksize;
3338         int ret = 0;
3339         int nr = 0;
3340         bool compressed;
3341
3342         if (tree->ops && tree->ops->writepage_start_hook) {
3343                 ret = tree->ops->writepage_start_hook(page, start,
3344                                                       page_end);
3345                 if (ret) {
3346                         /* Fixup worker will requeue */
3347                         if (ret == -EBUSY)
3348                                 wbc->pages_skipped++;
3349                         else
3350                                 redirty_page_for_writepage(wbc, page);
3351
3352                         update_nr_written(wbc, nr_written);
3353                         unlock_page(page);
3354                         return 1;
3355                 }
3356         }
3357
3358         /*
3359          * we don't want to touch the inode after unlocking the page,
3360          * so we update the mapping writeback index now
3361          */
3362         update_nr_written(wbc, nr_written + 1);
3363
3364         end = page_end;
3365         if (i_size <= start) {
3366                 if (tree->ops && tree->ops->writepage_end_io_hook)
3367                         tree->ops->writepage_end_io_hook(page, start,
3368                                                          page_end, NULL, 1);
3369                 goto done;
3370         }
3371
3372         blocksize = inode->i_sb->s_blocksize;
3373
3374         while (cur <= end) {
3375                 u64 em_end;
3376
3377                 if (cur >= i_size) {
3378                         if (tree->ops && tree->ops->writepage_end_io_hook)
3379                                 tree->ops->writepage_end_io_hook(page, cur,
3380                                                          page_end, NULL, 1);
3381                         break;
3382                 }
3383                 em = epd->get_extent(BTRFS_I(inode), page, pg_offset, cur,
3384                                      end - cur + 1, 1);
3385                 if (IS_ERR_OR_NULL(em)) {
3386                         SetPageError(page);
3387                         ret = PTR_ERR_OR_ZERO(em);
3388                         break;
3389                 }
3390
3391                 extent_offset = cur - em->start;
3392                 em_end = extent_map_end(em);
3393                 BUG_ON(em_end <= cur);
3394                 BUG_ON(end < cur);
3395                 iosize = min(em_end - cur, end - cur + 1);
3396                 iosize = ALIGN(iosize, blocksize);
3397                 sector = (em->block_start + extent_offset) >> 9;
3398                 bdev = em->bdev;
3399                 block_start = em->block_start;
3400                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3401                 free_extent_map(em);
3402                 em = NULL;
3403
3404                 /*
3405                  * compressed and inline extents are written through other
3406                  * paths in the FS
3407                  */
3408                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3409                     block_start == EXTENT_MAP_INLINE) {
3410                         /*
3411                          * end_io notification does not happen here for
3412                          * compressed extents
3413                          */
3414                         if (!compressed && tree->ops &&
3415                             tree->ops->writepage_end_io_hook)
3416                                 tree->ops->writepage_end_io_hook(page, cur,
3417                                                          cur + iosize - 1,
3418                                                          NULL, 1);
3419                         else if (compressed) {
3420                                 /* we don't want to end_page_writeback on
3421                                  * a compressed extent.  this happens
3422                                  * elsewhere
3423                                  */
3424                                 nr++;
3425                         }
3426
3427                         cur += iosize;
3428                         pg_offset += iosize;
3429                         continue;
3430                 }
3431
3432                 set_range_writeback(tree, cur, cur + iosize - 1);
3433                 if (!PageWriteback(page)) {
3434                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3435                                    "page %lu not writeback, cur %llu end %llu",
3436                                page->index, cur, end);
3437                 }
3438
3439                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3440                                          page, sector, iosize, pg_offset,
3441                                          bdev, &epd->bio,
3442                                          end_bio_extent_writepage,
3443                                          0, 0, 0, false);
3444                 if (ret) {
3445                         SetPageError(page);
3446                         if (PageWriteback(page))
3447                                 end_page_writeback(page);
3448                 }
3449
3450                 cur = cur + iosize;
3451                 pg_offset += iosize;
3452                 nr++;
3453         }
3454 done:
3455         *nr_ret = nr;
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_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_SHIFT;
3477         int write_flags = 0;
3478         unsigned long nr_written = 0;
3479
3480         if (wbc->sync_mode == WB_SYNC_ALL)
3481                 write_flags = REQ_SYNC;
3482
3483         trace___extent_writepage(page, inode, wbc);
3484
3485         WARN_ON(!PageLocked(page));
3486
3487         ClearPageError(page);
3488
3489         pg_offset = i_size & (PAGE_SIZE - 1);
3490         if (page->index > end_index ||
3491            (page->index == end_index && !pg_offset)) {
3492                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3493                 unlock_page(page);
3494                 return 0;
3495         }
3496
3497         if (page->index == end_index) {
3498                 char *userpage;
3499
3500                 userpage = kmap_atomic(page);
3501                 memset(userpage + pg_offset, 0,
3502                        PAGE_SIZE - pg_offset);
3503                 kunmap_atomic(userpage);
3504                 flush_dcache_page(page);
3505         }
3506
3507         pg_offset = 0;
3508
3509         set_page_extent_mapped(page);
3510
3511         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3512         if (ret == 1)
3513                 goto done_unlocked;
3514         if (ret)
3515                 goto done;
3516
3517         ret = __extent_writepage_io(inode, page, wbc, epd,
3518                                     i_size, nr_written, write_flags, &nr);
3519         if (ret == 1)
3520                 goto done_unlocked;
3521
3522 done:
3523         if (nr == 0) {
3524                 /* make sure the mapping tag for page dirty gets cleared */
3525                 set_page_writeback(page);
3526                 end_page_writeback(page);
3527         }
3528         if (PageError(page)) {
3529                 ret = ret < 0 ? ret : -EIO;
3530                 end_extent_writepage(page, ret, start, page_end);
3531         }
3532         unlock_page(page);
3533         return ret;
3534
3535 done_unlocked:
3536         return 0;
3537 }
3538
3539 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3540 {
3541         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3542                        TASK_UNINTERRUPTIBLE);
3543 }
3544
3545 static noinline_for_stack int
3546 lock_extent_buffer_for_io(struct extent_buffer *eb,
3547                           struct btrfs_fs_info *fs_info,
3548                           struct extent_page_data *epd)
3549 {
3550         unsigned long i, num_pages;
3551         int flush = 0;
3552         int ret = 0;
3553
3554         if (!btrfs_try_tree_write_lock(eb)) {
3555                 flush = 1;
3556                 flush_write_bio(epd);
3557                 btrfs_tree_lock(eb);
3558         }
3559
3560         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3561                 btrfs_tree_unlock(eb);
3562                 if (!epd->sync_io)
3563                         return 0;
3564                 if (!flush) {
3565                         flush_write_bio(epd);
3566                         flush = 1;
3567                 }
3568                 while (1) {
3569                         wait_on_extent_buffer_writeback(eb);
3570                         btrfs_tree_lock(eb);
3571                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3572                                 break;
3573                         btrfs_tree_unlock(eb);
3574                 }
3575         }
3576
3577         /*
3578          * We need to do this to prevent races in people who check if the eb is
3579          * under IO since we can end up having no IO bits set for a short period
3580          * of time.
3581          */
3582         spin_lock(&eb->refs_lock);
3583         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3584                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3585                 spin_unlock(&eb->refs_lock);
3586                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3587                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3588                                      -eb->len,
3589                                      fs_info->dirty_metadata_batch);
3590                 ret = 1;
3591         } else {
3592                 spin_unlock(&eb->refs_lock);
3593         }
3594
3595         btrfs_tree_unlock(eb);
3596
3597         if (!ret)
3598                 return ret;
3599
3600         num_pages = num_extent_pages(eb->start, eb->len);
3601         for (i = 0; i < num_pages; i++) {
3602                 struct page *p = eb->pages[i];
3603
3604                 if (!trylock_page(p)) {
3605                         if (!flush) {
3606                                 flush_write_bio(epd);
3607                                 flush = 1;
3608                         }
3609                         lock_page(p);
3610                 }
3611         }
3612
3613         return ret;
3614 }
3615
3616 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3617 {
3618         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3619         smp_mb__after_atomic();
3620         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3621 }
3622
3623 static void set_btree_ioerr(struct page *page)
3624 {
3625         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3626
3627         SetPageError(page);
3628         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3629                 return;
3630
3631         /*
3632          * If writeback for a btree extent that doesn't belong to a log tree
3633          * failed, increment the counter transaction->eb_write_errors.
3634          * We do this because while the transaction is running and before it's
3635          * committing (when we call filemap_fdata[write|wait]_range against
3636          * the btree inode), we might have
3637          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3638          * returns an error or an error happens during writeback, when we're
3639          * committing the transaction we wouldn't know about it, since the pages
3640          * can be no longer dirty nor marked anymore for writeback (if a
3641          * subsequent modification to the extent buffer didn't happen before the
3642          * transaction commit), which makes filemap_fdata[write|wait]_range not
3643          * able to find the pages tagged with SetPageError at transaction
3644          * commit time. So if this happens we must abort the transaction,
3645          * otherwise we commit a super block with btree roots that point to
3646          * btree nodes/leafs whose content on disk is invalid - either garbage
3647          * or the content of some node/leaf from a past generation that got
3648          * cowed or deleted and is no longer valid.
3649          *
3650          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3651          * not be enough - we need to distinguish between log tree extents vs
3652          * non-log tree extents, and the next filemap_fdatawait_range() call
3653          * will catch and clear such errors in the mapping - and that call might
3654          * be from a log sync and not from a transaction commit. Also, checking
3655          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3656          * not done and would not be reliable - the eb might have been released
3657          * from memory and reading it back again means that flag would not be
3658          * set (since it's a runtime flag, not persisted on disk).
3659          *
3660          * Using the flags below in the btree inode also makes us achieve the
3661          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3662          * writeback for all dirty pages and before filemap_fdatawait_range()
3663          * is called, the writeback for all dirty pages had already finished
3664          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3665          * filemap_fdatawait_range() would return success, as it could not know
3666          * that writeback errors happened (the pages were no longer tagged for
3667          * writeback).
3668          */
3669         switch (eb->log_index) {
3670         case -1:
3671                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3672                 break;
3673         case 0:
3674                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3675                 break;
3676         case 1:
3677                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3678                 break;
3679         default:
3680                 BUG(); /* unexpected, logic error */
3681         }
3682 }
3683
3684 static void end_bio_extent_buffer_writepage(struct bio *bio)
3685 {
3686         struct bio_vec *bvec;
3687         struct extent_buffer *eb;
3688         int i, done;
3689
3690         bio_for_each_segment_all(bvec, bio, i) {
3691                 struct page *page = bvec->bv_page;
3692
3693                 eb = (struct extent_buffer *)page->private;
3694                 BUG_ON(!eb);
3695                 done = atomic_dec_and_test(&eb->io_pages);
3696
3697                 if (bio->bi_error ||
3698                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3699                         ClearPageUptodate(page);
3700                         set_btree_ioerr(page);
3701                 }
3702
3703                 end_page_writeback(page);
3704
3705                 if (!done)
3706                         continue;
3707
3708                 end_extent_buffer_writeback(eb);
3709         }
3710
3711         bio_put(bio);
3712 }
3713
3714 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3715                         struct btrfs_fs_info *fs_info,
3716                         struct writeback_control *wbc,
3717                         struct extent_page_data *epd)
3718 {
3719         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3720         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3721         u64 offset = eb->start;
3722         u32 nritems;
3723         unsigned long i, num_pages;
3724         unsigned long bio_flags = 0;
3725         unsigned long start, end;
3726         int write_flags = (epd->sync_io ? REQ_SYNC : 0) | REQ_META;
3727         int ret = 0;
3728
3729         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3730         num_pages = num_extent_pages(eb->start, eb->len);
3731         atomic_set(&eb->io_pages, num_pages);
3732         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3733                 bio_flags = EXTENT_BIO_TREE_LOG;
3734
3735         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3736         nritems = btrfs_header_nritems(eb);
3737         if (btrfs_header_level(eb) > 0) {
3738                 end = btrfs_node_key_ptr_offset(nritems);
3739
3740                 memzero_extent_buffer(eb, end, eb->len - end);
3741         } else {
3742                 /*
3743                  * leaf:
3744                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3745                  */
3746                 start = btrfs_item_nr_offset(nritems);
3747                 end = btrfs_leaf_data(eb) + leaf_data_end(fs_info, eb);
3748                 memzero_extent_buffer(eb, start, end - start);
3749         }
3750
3751         for (i = 0; i < num_pages; i++) {
3752                 struct page *p = eb->pages[i];
3753
3754                 clear_page_dirty_for_io(p);
3755                 set_page_writeback(p);
3756                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3757                                          p, offset >> 9, PAGE_SIZE, 0, bdev,
3758                                          &epd->bio,
3759                                          end_bio_extent_buffer_writepage,
3760                                          0, epd->bio_flags, bio_flags, false);
3761                 epd->bio_flags = bio_flags;
3762                 if (ret) {
3763                         set_btree_ioerr(p);
3764                         if (PageWriteback(p))
3765                                 end_page_writeback(p);
3766                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3767                                 end_extent_buffer_writeback(eb);
3768                         ret = -EIO;
3769                         break;
3770                 }
3771                 offset += PAGE_SIZE;
3772                 update_nr_written(wbc, 1);
3773                 unlock_page(p);
3774         }
3775
3776         if (unlikely(ret)) {
3777                 for (; i < num_pages; i++) {
3778                         struct page *p = eb->pages[i];
3779                         clear_page_dirty_for_io(p);
3780                         unlock_page(p);
3781                 }
3782         }
3783
3784         return ret;
3785 }
3786
3787 int btree_write_cache_pages(struct address_space *mapping,
3788                                    struct writeback_control *wbc)
3789 {
3790         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3791         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3792         struct extent_buffer *eb, *prev_eb = NULL;
3793         struct extent_page_data epd = {
3794                 .bio = NULL,
3795                 .tree = tree,
3796                 .extent_locked = 0,
3797                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3798                 .bio_flags = 0,
3799         };
3800         int ret = 0;
3801         int done = 0;
3802         int nr_to_write_done = 0;
3803         struct pagevec pvec;
3804         int nr_pages;
3805         pgoff_t index;
3806         pgoff_t end;            /* Inclusive */
3807         int scanned = 0;
3808         int tag;
3809
3810         pagevec_init(&pvec, 0);
3811         if (wbc->range_cyclic) {
3812                 index = mapping->writeback_index; /* Start from prev offset */
3813                 end = -1;
3814         } else {
3815                 index = wbc->range_start >> PAGE_SHIFT;
3816                 end = wbc->range_end >> PAGE_SHIFT;
3817                 scanned = 1;
3818         }
3819         if (wbc->sync_mode == WB_SYNC_ALL)
3820                 tag = PAGECACHE_TAG_TOWRITE;
3821         else
3822                 tag = PAGECACHE_TAG_DIRTY;
3823 retry:
3824         if (wbc->sync_mode == WB_SYNC_ALL)
3825                 tag_pages_for_writeback(mapping, index, end);
3826         while (!done && !nr_to_write_done && (index <= end) &&
3827                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3828                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3829                 unsigned i;
3830
3831                 scanned = 1;
3832                 for (i = 0; i < nr_pages; i++) {
3833                         struct page *page = pvec.pages[i];
3834
3835                         if (!PagePrivate(page))
3836                                 continue;
3837
3838                         if (!wbc->range_cyclic && page->index > end) {
3839                                 done = 1;
3840                                 break;
3841                         }
3842
3843                         spin_lock(&mapping->private_lock);
3844                         if (!PagePrivate(page)) {
3845                                 spin_unlock(&mapping->private_lock);
3846                                 continue;
3847                         }
3848
3849                         eb = (struct extent_buffer *)page->private;
3850
3851                         /*
3852                          * Shouldn't happen and normally this would be a BUG_ON
3853                          * but no sense in crashing the users box for something
3854                          * we can survive anyway.
3855                          */
3856                         if (WARN_ON(!eb)) {
3857                                 spin_unlock(&mapping->private_lock);
3858                                 continue;
3859                         }
3860
3861                         if (eb == prev_eb) {
3862                                 spin_unlock(&mapping->private_lock);
3863                                 continue;
3864                         }
3865
3866                         ret = atomic_inc_not_zero(&eb->refs);
3867                         spin_unlock(&mapping->private_lock);
3868                         if (!ret)
3869                                 continue;
3870
3871                         prev_eb = eb;
3872                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3873                         if (!ret) {
3874                                 free_extent_buffer(eb);
3875                                 continue;
3876                         }
3877
3878                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3879                         if (ret) {
3880                                 done = 1;
3881                                 free_extent_buffer(eb);
3882                                 break;
3883                         }
3884                         free_extent_buffer(eb);
3885
3886                         /*
3887                          * the filesystem may choose to bump up nr_to_write.
3888                          * We have to make sure to honor the new nr_to_write
3889                          * at any time
3890                          */
3891                         nr_to_write_done = wbc->nr_to_write <= 0;
3892                 }
3893                 pagevec_release(&pvec);
3894                 cond_resched();
3895         }
3896         if (!scanned && !done) {
3897                 /*
3898                  * We hit the last page and there is more work to be done: wrap
3899                  * back to the start of the file
3900                  */
3901                 scanned = 1;
3902                 index = 0;
3903                 goto retry;
3904         }
3905         flush_write_bio(&epd);
3906         return ret;
3907 }
3908
3909 /**
3910  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3911  * @mapping: address space structure to write
3912  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3913  * @writepage: function called for each page
3914  * @data: data passed to writepage function
3915  *
3916  * If a page is already under I/O, write_cache_pages() skips it, even
3917  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3918  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3919  * and msync() need to guarantee that all the data which was dirty at the time
3920  * the call was made get new I/O started against them.  If wbc->sync_mode is
3921  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3922  * existing IO to complete.
3923  */
3924 static int extent_write_cache_pages(struct address_space *mapping,
3925                              struct writeback_control *wbc,
3926                              writepage_t writepage, void *data,
3927                              void (*flush_fn)(void *))
3928 {
3929         struct inode *inode = mapping->host;
3930         int ret = 0;
3931         int done = 0;
3932         int nr_to_write_done = 0;
3933         struct pagevec pvec;
3934         int nr_pages;
3935         pgoff_t index;
3936         pgoff_t end;            /* Inclusive */
3937         pgoff_t done_index;
3938         int range_whole = 0;
3939         int scanned = 0;
3940         int tag;
3941
3942         /*
3943          * We have to hold onto the inode so that ordered extents can do their
3944          * work when the IO finishes.  The alternative to this is failing to add
3945          * an ordered extent if the igrab() fails there and that is a huge pain
3946          * to deal with, so instead just hold onto the inode throughout the
3947          * writepages operation.  If it fails here we are freeing up the inode
3948          * anyway and we'd rather not waste our time writing out stuff that is
3949          * going to be truncated anyway.
3950          */
3951         if (!igrab(inode))
3952                 return 0;
3953
3954         pagevec_init(&pvec, 0);
3955         if (wbc->range_cyclic) {
3956                 index = mapping->writeback_index; /* Start from prev offset */
3957                 end = -1;
3958         } else {
3959                 index = wbc->range_start >> PAGE_SHIFT;
3960                 end = wbc->range_end >> PAGE_SHIFT;
3961                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3962                         range_whole = 1;
3963                 scanned = 1;
3964         }
3965         if (wbc->sync_mode == WB_SYNC_ALL)
3966                 tag = PAGECACHE_TAG_TOWRITE;
3967         else
3968                 tag = PAGECACHE_TAG_DIRTY;
3969 retry:
3970         if (wbc->sync_mode == WB_SYNC_ALL)
3971                 tag_pages_for_writeback(mapping, index, end);
3972         done_index = index;
3973         while (!done && !nr_to_write_done && (index <= end) &&
3974                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3975                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3976                 unsigned i;
3977
3978                 scanned = 1;
3979                 for (i = 0; i < nr_pages; i++) {
3980                         struct page *page = pvec.pages[i];
3981
3982                         done_index = page->index;
3983                         /*
3984                          * At this point we hold neither mapping->tree_lock nor
3985                          * lock on the page itself: the page may be truncated or
3986                          * invalidated (changing page->mapping to NULL), or even
3987                          * swizzled back from swapper_space to tmpfs file
3988                          * mapping
3989                          */
3990                         if (!trylock_page(page)) {
3991                                 flush_fn(data);
3992                                 lock_page(page);
3993                         }
3994
3995                         if (unlikely(page->mapping != mapping)) {
3996                                 unlock_page(page);
3997                                 continue;
3998                         }
3999
4000                         if (!wbc->range_cyclic && page->index > end) {
4001                                 done = 1;
4002                                 unlock_page(page);
4003                                 continue;
4004                         }
4005
4006                         if (wbc->sync_mode != WB_SYNC_NONE) {
4007                                 if (PageWriteback(page))
4008                                         flush_fn(data);
4009                                 wait_on_page_writeback(page);
4010                         }
4011
4012                         if (PageWriteback(page) ||
4013                             !clear_page_dirty_for_io(page)) {
4014                                 unlock_page(page);
4015                                 continue;
4016                         }
4017
4018                         ret = (*writepage)(page, wbc, data);
4019
4020                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4021                                 unlock_page(page);
4022                                 ret = 0;
4023                         }
4024                         if (ret < 0) {
4025                                 /*
4026                                  * done_index is set past this page,
4027                                  * so media errors will not choke
4028                                  * background writeout for the entire
4029                                  * file. This has consequences for
4030                                  * range_cyclic semantics (ie. it may
4031                                  * not be suitable for data integrity
4032                                  * writeout).
4033                                  */
4034                                 done_index = page->index + 1;
4035                                 done = 1;
4036                                 break;
4037                         }
4038
4039                         /*
4040                          * the filesystem may choose to bump up nr_to_write.
4041                          * We have to make sure to honor the new nr_to_write
4042                          * at any time
4043                          */
4044                         nr_to_write_done = wbc->nr_to_write <= 0;
4045                 }
4046                 pagevec_release(&pvec);
4047                 cond_resched();
4048         }
4049         if (!scanned && !done) {
4050                 /*
4051                  * We hit the last page and there is more work to be done: wrap
4052                  * back to the start of the file
4053                  */
4054                 scanned = 1;
4055                 index = 0;
4056                 goto retry;
4057         }
4058
4059         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4060                 mapping->writeback_index = done_index;
4061
4062         btrfs_add_delayed_iput(inode);
4063         return ret;
4064 }
4065
4066 static void flush_epd_write_bio(struct extent_page_data *epd)
4067 {
4068         if (epd->bio) {
4069                 int ret;
4070
4071                 bio_set_op_attrs(epd->bio, REQ_OP_WRITE,
4072                                  epd->sync_io ? REQ_SYNC : 0);
4073
4074                 ret = submit_one_bio(epd->bio, 0, epd->bio_flags);
4075                 BUG_ON(ret < 0); /* -ENOMEM */
4076                 epd->bio = NULL;
4077         }
4078 }
4079
4080 static noinline void flush_write_bio(void *data)
4081 {
4082         struct extent_page_data *epd = data;
4083         flush_epd_write_bio(epd);
4084 }
4085
4086 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4087                           get_extent_t *get_extent,
4088                           struct writeback_control *wbc)
4089 {
4090         int ret;
4091         struct extent_page_data epd = {
4092                 .bio = NULL,
4093                 .tree = tree,
4094                 .get_extent = get_extent,
4095                 .extent_locked = 0,
4096                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4097                 .bio_flags = 0,
4098         };
4099
4100         ret = __extent_writepage(page, wbc, &epd);
4101
4102         flush_epd_write_bio(&epd);
4103         return ret;
4104 }
4105
4106 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4107                               u64 start, u64 end, get_extent_t *get_extent,
4108                               int mode)
4109 {
4110         int ret = 0;
4111         struct address_space *mapping = inode->i_mapping;
4112         struct page *page;
4113         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4114                 PAGE_SHIFT;
4115
4116         struct extent_page_data epd = {
4117                 .bio = NULL,
4118                 .tree = tree,
4119                 .get_extent = get_extent,
4120                 .extent_locked = 1,
4121                 .sync_io = mode == WB_SYNC_ALL,
4122                 .bio_flags = 0,
4123         };
4124         struct writeback_control wbc_writepages = {
4125                 .sync_mode      = mode,
4126                 .nr_to_write    = nr_pages * 2,
4127                 .range_start    = start,
4128                 .range_end      = end + 1,
4129         };
4130
4131         while (start <= end) {
4132                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4133                 if (clear_page_dirty_for_io(page))
4134                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4135                 else {
4136                         if (tree->ops && tree->ops->writepage_end_io_hook)
4137                                 tree->ops->writepage_end_io_hook(page, start,
4138                                                  start + PAGE_SIZE - 1,
4139                                                  NULL, 1);
4140                         unlock_page(page);
4141                 }
4142                 put_page(page);
4143                 start += PAGE_SIZE;
4144         }
4145
4146         flush_epd_write_bio(&epd);
4147         return ret;
4148 }
4149
4150 int extent_writepages(struct extent_io_tree *tree,
4151                       struct address_space *mapping,
4152                       get_extent_t *get_extent,
4153                       struct writeback_control *wbc)
4154 {
4155         int ret = 0;
4156         struct extent_page_data epd = {
4157                 .bio = NULL,
4158                 .tree = tree,
4159                 .get_extent = get_extent,
4160                 .extent_locked = 0,
4161                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4162                 .bio_flags = 0,
4163         };
4164
4165         ret = extent_write_cache_pages(mapping, wbc, __extent_writepage, &epd,
4166                                        flush_write_bio);
4167         flush_epd_write_bio(&epd);
4168         return ret;
4169 }
4170
4171 int extent_readpages(struct extent_io_tree *tree,
4172                      struct address_space *mapping,
4173                      struct list_head *pages, unsigned nr_pages,
4174                      get_extent_t get_extent)
4175 {
4176         struct bio *bio = NULL;
4177         unsigned page_idx;
4178         unsigned long bio_flags = 0;
4179         struct page *pagepool[16];
4180         struct page *page;
4181         struct extent_map *em_cached = NULL;
4182         int nr = 0;
4183         u64 prev_em_start = (u64)-1;
4184
4185         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4186                 page = list_entry(pages->prev, struct page, lru);
4187
4188                 prefetchw(&page->flags);
4189                 list_del(&page->lru);
4190                 if (add_to_page_cache_lru(page, mapping,
4191                                         page->index,
4192                                         readahead_gfp_mask(mapping))) {
4193                         put_page(page);
4194                         continue;
4195                 }
4196
4197                 pagepool[nr++] = page;
4198                 if (nr < ARRAY_SIZE(pagepool))
4199                         continue;
4200                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4201                                    &bio, 0, &bio_flags, &prev_em_start);
4202                 nr = 0;
4203         }
4204         if (nr)
4205                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4206                                    &bio, 0, &bio_flags, &prev_em_start);
4207
4208         if (em_cached)
4209                 free_extent_map(em_cached);
4210
4211         BUG_ON(!list_empty(pages));
4212         if (bio)
4213                 return submit_one_bio(bio, 0, bio_flags);
4214         return 0;
4215 }
4216
4217 /*
4218  * basic invalidatepage code, this waits on any locked or writeback
4219  * ranges corresponding to the page, and then deletes any extent state
4220  * records from the tree
4221  */
4222 int extent_invalidatepage(struct extent_io_tree *tree,
4223                           struct page *page, unsigned long offset)
4224 {
4225         struct extent_state *cached_state = NULL;
4226         u64 start = page_offset(page);
4227         u64 end = start + PAGE_SIZE - 1;
4228         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4229
4230         start += ALIGN(offset, blocksize);
4231         if (start > end)
4232                 return 0;
4233
4234         lock_extent_bits(tree, start, end, &cached_state);
4235         wait_on_page_writeback(page);
4236         clear_extent_bit(tree, start, end,
4237                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4238                          EXTENT_DO_ACCOUNTING,
4239                          1, 1, &cached_state, GFP_NOFS);
4240         return 0;
4241 }
4242
4243 /*
4244  * a helper for releasepage, this tests for areas of the page that
4245  * are locked or under IO and drops the related state bits if it is safe
4246  * to drop the page.
4247  */
4248 static int try_release_extent_state(struct extent_map_tree *map,
4249                                     struct extent_io_tree *tree,
4250                                     struct page *page, gfp_t mask)
4251 {
4252         u64 start = page_offset(page);
4253         u64 end = start + PAGE_SIZE - 1;
4254         int ret = 1;
4255
4256         if (test_range_bit(tree, start, end,
4257                            EXTENT_IOBITS, 0, NULL))
4258                 ret = 0;
4259         else {
4260                 /*
4261                  * at this point we can safely clear everything except the
4262                  * locked bit and the nodatasum bit
4263                  */
4264                 ret = clear_extent_bit(tree, start, end,
4265                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4266                                  0, 0, NULL, mask);
4267
4268                 /* if clear_extent_bit failed for enomem reasons,
4269                  * we can't allow the release to continue.
4270                  */
4271                 if (ret < 0)
4272                         ret = 0;
4273                 else
4274                         ret = 1;
4275         }
4276         return ret;
4277 }
4278
4279 /*
4280  * a helper for releasepage.  As long as there are no locked extents
4281  * in the range corresponding to the page, both state records and extent
4282  * map records are removed
4283  */
4284 int try_release_extent_mapping(struct extent_map_tree *map,
4285                                struct extent_io_tree *tree, struct page *page,
4286                                gfp_t mask)
4287 {
4288         struct extent_map *em;
4289         u64 start = page_offset(page);
4290         u64 end = start + PAGE_SIZE - 1;
4291
4292         if (gfpflags_allow_blocking(mask) &&
4293             page->mapping->host->i_size > SZ_16M) {
4294                 u64 len;
4295                 while (start <= end) {
4296                         len = end - start + 1;
4297                         write_lock(&map->lock);
4298                         em = lookup_extent_mapping(map, start, len);
4299                         if (!em) {
4300                                 write_unlock(&map->lock);
4301                                 break;
4302                         }
4303                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4304                             em->start != start) {
4305                                 write_unlock(&map->lock);
4306                                 free_extent_map(em);
4307                                 break;
4308                         }
4309                         if (!test_range_bit(tree, em->start,
4310                                             extent_map_end(em) - 1,
4311                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4312                                             0, NULL)) {
4313                                 remove_extent_mapping(map, em);
4314                                 /* once for the rb tree */
4315                                 free_extent_map(em);
4316                         }
4317                         start = extent_map_end(em);
4318                         write_unlock(&map->lock);
4319
4320                         /* once for us */
4321                         free_extent_map(em);
4322                 }
4323         }
4324         return try_release_extent_state(map, tree, page, mask);
4325 }
4326
4327 /*
4328  * helper function for fiemap, which doesn't want to see any holes.
4329  * This maps until we find something past 'last'
4330  */
4331 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4332                                                 u64 offset,
4333                                                 u64 last,
4334                                                 get_extent_t *get_extent)
4335 {
4336         u64 sectorsize = btrfs_inode_sectorsize(inode);
4337         struct extent_map *em;
4338         u64 len;
4339
4340         if (offset >= last)
4341                 return NULL;
4342
4343         while (1) {
4344                 len = last - offset;
4345                 if (len == 0)
4346                         break;
4347                 len = ALIGN(len, sectorsize);
4348                 em = get_extent(BTRFS_I(inode), NULL, 0, offset, len, 0);
4349                 if (IS_ERR_OR_NULL(em))
4350                         return em;
4351
4352                 /* if this isn't a hole return it */
4353                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4354                     em->block_start != EXTENT_MAP_HOLE) {
4355                         return em;
4356                 }
4357
4358                 /* this is a hole, advance to the next extent */
4359                 offset = extent_map_end(em);
4360                 free_extent_map(em);
4361                 if (offset >= last)
4362                         break;
4363         }
4364         return NULL;
4365 }
4366
4367 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4368                 __u64 start, __u64 len, get_extent_t *get_extent)
4369 {
4370         int ret = 0;
4371         u64 off = start;
4372         u64 max = start + len;
4373         u32 flags = 0;
4374         u32 found_type;
4375         u64 last;
4376         u64 last_for_get_extent = 0;
4377         u64 disko = 0;
4378         u64 isize = i_size_read(inode);
4379         struct btrfs_key found_key;
4380         struct extent_map *em = NULL;
4381         struct extent_state *cached_state = NULL;
4382         struct btrfs_path *path;
4383         struct btrfs_root *root = BTRFS_I(inode)->root;
4384         int end = 0;
4385         u64 em_start = 0;
4386         u64 em_len = 0;
4387         u64 em_end = 0;
4388
4389         if (len == 0)
4390                 return -EINVAL;
4391
4392         path = btrfs_alloc_path();
4393         if (!path)
4394                 return -ENOMEM;
4395         path->leave_spinning = 1;
4396
4397         start = round_down(start, btrfs_inode_sectorsize(inode));
4398         len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
4399
4400         /*
4401          * lookup the last file extent.  We're not using i_size here
4402          * because there might be preallocation past i_size
4403          */
4404         ret = btrfs_lookup_file_extent(NULL, root, path,
4405                         btrfs_ino(BTRFS_I(inode)), -1, 0);
4406         if (ret < 0) {
4407                 btrfs_free_path(path);
4408                 return ret;
4409         } else {
4410                 WARN_ON(!ret);
4411                 if (ret == 1)
4412                         ret = 0;
4413         }
4414
4415         path->slots[0]--;
4416         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4417         found_type = found_key.type;
4418
4419         /* No extents, but there might be delalloc bits */
4420         if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) ||
4421             found_type != BTRFS_EXTENT_DATA_KEY) {
4422                 /* have to trust i_size as the end */
4423                 last = (u64)-1;
4424                 last_for_get_extent = isize;
4425         } else {
4426                 /*
4427                  * remember the start of the last extent.  There are a
4428                  * bunch of different factors that go into the length of the
4429                  * extent, so its much less complex to remember where it started
4430                  */
4431                 last = found_key.offset;
4432                 last_for_get_extent = last + 1;
4433         }
4434         btrfs_release_path(path);
4435
4436         /*
4437          * we might have some extents allocated but more delalloc past those
4438          * extents.  so, we trust isize unless the start of the last extent is
4439          * beyond isize
4440          */
4441         if (last < isize) {
4442                 last = (u64)-1;
4443                 last_for_get_extent = isize;
4444         }
4445
4446         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4447                          &cached_state);
4448
4449         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4450                                    get_extent);
4451         if (!em)
4452                 goto out;
4453         if (IS_ERR(em)) {
4454                 ret = PTR_ERR(em);
4455                 goto out;
4456         }
4457
4458         while (!end) {
4459                 u64 offset_in_extent = 0;
4460
4461                 /* break if the extent we found is outside the range */
4462                 if (em->start >= max || extent_map_end(em) < off)
4463                         break;
4464
4465                 /*
4466                  * get_extent may return an extent that starts before our
4467                  * requested range.  We have to make sure the ranges
4468                  * we return to fiemap always move forward and don't
4469                  * overlap, so adjust the offsets here
4470                  */
4471                 em_start = max(em->start, off);
4472
4473                 /*
4474                  * record the offset from the start of the extent
4475                  * for adjusting the disk offset below.  Only do this if the
4476                  * extent isn't compressed since our in ram offset may be past
4477                  * what we have actually allocated on disk.
4478                  */
4479                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4480                         offset_in_extent = em_start - em->start;
4481                 em_end = extent_map_end(em);
4482                 em_len = em_end - em_start;
4483                 disko = 0;
4484                 flags = 0;
4485
4486                 /*
4487                  * bump off for our next call to get_extent
4488                  */
4489                 off = extent_map_end(em);
4490                 if (off >= max)
4491                         end = 1;
4492
4493                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4494                         end = 1;
4495                         flags |= FIEMAP_EXTENT_LAST;
4496                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4497                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4498                                   FIEMAP_EXTENT_NOT_ALIGNED);
4499                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4500                         flags |= (FIEMAP_EXTENT_DELALLOC |
4501                                   FIEMAP_EXTENT_UNKNOWN);
4502                 } else if (fieinfo->fi_extents_max) {
4503                         struct btrfs_trans_handle *trans;
4504
4505                         u64 bytenr = em->block_start -
4506                                 (em->start - em->orig_start);
4507
4508                         disko = em->block_start + offset_in_extent;
4509
4510                         /*
4511                          * We need a trans handle to get delayed refs
4512                          */
4513                         trans = btrfs_join_transaction(root);
4514                         /*
4515                          * It's OK if we can't start a trans we can still check
4516                          * from commit_root
4517                          */
4518                         if (IS_ERR(trans))
4519                                 trans = NULL;
4520
4521                         /*
4522                          * As btrfs supports shared space, this information
4523                          * can be exported to userspace tools via
4524                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4525                          * then we're just getting a count and we can skip the
4526                          * lookup stuff.
4527                          */
4528                         ret = btrfs_check_shared(trans, root->fs_info,
4529                                         root->objectid,
4530                                         btrfs_ino(BTRFS_I(inode)), bytenr);
4531                         if (trans)
4532                                 btrfs_end_transaction(trans);
4533                         if (ret < 0)
4534                                 goto out_free;
4535                         if (ret)
4536                                 flags |= FIEMAP_EXTENT_SHARED;
4537                         ret = 0;
4538                 }
4539                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4540                         flags |= FIEMAP_EXTENT_ENCODED;
4541                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4542                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4543
4544                 free_extent_map(em);
4545                 em = NULL;
4546                 if ((em_start >= last) || em_len == (u64)-1 ||
4547                    (last == (u64)-1 && isize <= em_end)) {
4548                         flags |= FIEMAP_EXTENT_LAST;
4549                         end = 1;
4550                 }
4551
4552                 /* now scan forward to see if this is really the last extent. */
4553                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4554                                            get_extent);
4555                 if (IS_ERR(em)) {
4556                         ret = PTR_ERR(em);
4557                         goto out;
4558                 }
4559                 if (!em) {
4560                         flags |= FIEMAP_EXTENT_LAST;
4561                         end = 1;
4562                 }
4563                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4564                                               em_len, flags);
4565                 if (ret) {
4566                         if (ret == 1)
4567                                 ret = 0;
4568                         goto out_free;
4569                 }
4570         }
4571 out_free:
4572         free_extent_map(em);
4573 out:
4574         btrfs_free_path(path);
4575         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4576                              &cached_state, GFP_NOFS);
4577         return ret;
4578 }
4579
4580 static void __free_extent_buffer(struct extent_buffer *eb)
4581 {
4582         btrfs_leak_debug_del(&eb->leak_list);
4583         kmem_cache_free(extent_buffer_cache, eb);
4584 }
4585
4586 int extent_buffer_under_io(struct extent_buffer *eb)
4587 {
4588         return (atomic_read(&eb->io_pages) ||
4589                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4590                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4591 }
4592
4593 /*
4594  * Helper for releasing extent buffer page.
4595  */
4596 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4597 {
4598         unsigned long index;
4599         struct page *page;
4600         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4601
4602         BUG_ON(extent_buffer_under_io(eb));
4603
4604         index = num_extent_pages(eb->start, eb->len);
4605         if (index == 0)
4606                 return;
4607
4608         do {
4609                 index--;
4610                 page = eb->pages[index];
4611                 if (!page)
4612                         continue;
4613                 if (mapped)
4614                         spin_lock(&page->mapping->private_lock);
4615                 /*
4616                  * We do this since we'll remove the pages after we've
4617                  * removed the eb from the radix tree, so we could race
4618                  * and have this page now attached to the new eb.  So
4619                  * only clear page_private if it's still connected to
4620                  * this eb.
4621                  */
4622                 if (PagePrivate(page) &&
4623                     page->private == (unsigned long)eb) {
4624                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4625                         BUG_ON(PageDirty(page));
4626                         BUG_ON(PageWriteback(page));
4627                         /*
4628                          * We need to make sure we haven't be attached
4629                          * to a new eb.
4630                          */
4631                         ClearPagePrivate(page);
4632                         set_page_private(page, 0);
4633                         /* One for the page private */
4634                         put_page(page);
4635                 }
4636
4637                 if (mapped)
4638                         spin_unlock(&page->mapping->private_lock);
4639
4640                 /* One for when we allocated the page */
4641                 put_page(page);
4642         } while (index != 0);
4643 }
4644
4645 /*
4646  * Helper for releasing the extent buffer.
4647  */
4648 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4649 {
4650         btrfs_release_extent_buffer_page(eb);
4651         __free_extent_buffer(eb);
4652 }
4653
4654 static struct extent_buffer *
4655 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4656                       unsigned long len)
4657 {
4658         struct extent_buffer *eb = NULL;
4659
4660         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4661         eb->start = start;
4662         eb->len = len;
4663         eb->fs_info = fs_info;
4664         eb->bflags = 0;
4665         rwlock_init(&eb->lock);
4666         atomic_set(&eb->write_locks, 0);
4667         atomic_set(&eb->read_locks, 0);
4668         atomic_set(&eb->blocking_readers, 0);
4669         atomic_set(&eb->blocking_writers, 0);
4670         atomic_set(&eb->spinning_readers, 0);
4671         atomic_set(&eb->spinning_writers, 0);
4672         eb->lock_nested = 0;
4673         init_waitqueue_head(&eb->write_lock_wq);
4674         init_waitqueue_head(&eb->read_lock_wq);
4675
4676         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4677
4678         spin_lock_init(&eb->refs_lock);
4679         atomic_set(&eb->refs, 1);
4680         atomic_set(&eb->io_pages, 0);
4681
4682         /*
4683          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4684          */
4685         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4686                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4687         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4688
4689         return eb;
4690 }
4691
4692 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4693 {
4694         unsigned long i;
4695         struct page *p;
4696         struct extent_buffer *new;
4697         unsigned long num_pages = num_extent_pages(src->start, src->len);
4698
4699         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4700         if (new == NULL)
4701                 return NULL;
4702
4703         for (i = 0; i < num_pages; i++) {
4704                 p = alloc_page(GFP_NOFS);
4705                 if (!p) {
4706                         btrfs_release_extent_buffer(new);
4707                         return NULL;
4708                 }
4709                 attach_extent_buffer_page(new, p);
4710                 WARN_ON(PageDirty(p));
4711                 SetPageUptodate(p);
4712                 new->pages[i] = p;
4713                 copy_page(page_address(p), page_address(src->pages[i]));
4714         }
4715
4716         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4717         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4718
4719         return new;
4720 }
4721
4722 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4723                                                   u64 start, unsigned long len)
4724 {
4725         struct extent_buffer *eb;
4726         unsigned long num_pages;
4727         unsigned long i;
4728
4729         num_pages = num_extent_pages(start, len);
4730
4731         eb = __alloc_extent_buffer(fs_info, start, len);
4732         if (!eb)
4733                 return NULL;
4734
4735         for (i = 0; i < num_pages; i++) {
4736                 eb->pages[i] = alloc_page(GFP_NOFS);
4737                 if (!eb->pages[i])
4738                         goto err;
4739         }
4740         set_extent_buffer_uptodate(eb);
4741         btrfs_set_header_nritems(eb, 0);
4742         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4743
4744         return eb;
4745 err:
4746         for (; i > 0; i--)
4747                 __free_page(eb->pages[i - 1]);
4748         __free_extent_buffer(eb);
4749         return NULL;
4750 }
4751
4752 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4753                                                 u64 start)
4754 {
4755         return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
4756 }
4757
4758 static void check_buffer_tree_ref(struct extent_buffer *eb)
4759 {
4760         int refs;
4761         /* the ref bit is tricky.  We have to make sure it is set
4762          * if we have the buffer dirty.   Otherwise the
4763          * code to free a buffer can end up dropping a dirty
4764          * page
4765          *
4766          * Once the ref bit is set, it won't go away while the
4767          * buffer is dirty or in writeback, and it also won't
4768          * go away while we have the reference count on the
4769          * eb bumped.
4770          *
4771          * We can't just set the ref bit without bumping the
4772          * ref on the eb because free_extent_buffer might
4773          * see the ref bit and try to clear it.  If this happens
4774          * free_extent_buffer might end up dropping our original
4775          * ref by mistake and freeing the page before we are able
4776          * to add one more ref.
4777          *
4778          * So bump the ref count first, then set the bit.  If someone
4779          * beat us to it, drop the ref we added.
4780          */
4781         refs = atomic_read(&eb->refs);
4782         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4783                 return;
4784
4785         spin_lock(&eb->refs_lock);
4786         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4787                 atomic_inc(&eb->refs);
4788         spin_unlock(&eb->refs_lock);
4789 }
4790
4791 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4792                 struct page *accessed)
4793 {
4794         unsigned long num_pages, i;
4795
4796         check_buffer_tree_ref(eb);
4797
4798         num_pages = num_extent_pages(eb->start, eb->len);
4799         for (i = 0; i < num_pages; i++) {
4800                 struct page *p = eb->pages[i];
4801
4802                 if (p != accessed)
4803                         mark_page_accessed(p);
4804         }
4805 }
4806
4807 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4808                                          u64 start)
4809 {
4810         struct extent_buffer *eb;
4811
4812         rcu_read_lock();
4813         eb = radix_tree_lookup(&fs_info->buffer_radix,
4814                                start >> PAGE_SHIFT);
4815         if (eb && atomic_inc_not_zero(&eb->refs)) {
4816                 rcu_read_unlock();
4817                 /*
4818                  * Lock our eb's refs_lock to avoid races with
4819                  * free_extent_buffer. When we get our eb it might be flagged
4820                  * with EXTENT_BUFFER_STALE and another task running
4821                  * free_extent_buffer might have seen that flag set,
4822                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4823                  * writeback flags not set) and it's still in the tree (flag
4824                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4825                  * of decrementing the extent buffer's reference count twice.
4826                  * So here we could race and increment the eb's reference count,
4827                  * clear its stale flag, mark it as dirty and drop our reference
4828                  * before the other task finishes executing free_extent_buffer,
4829                  * which would later result in an attempt to free an extent
4830                  * buffer that is dirty.
4831                  */
4832                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4833                         spin_lock(&eb->refs_lock);
4834                         spin_unlock(&eb->refs_lock);
4835                 }
4836                 mark_extent_buffer_accessed(eb, NULL);
4837                 return eb;
4838         }
4839         rcu_read_unlock();
4840
4841         return NULL;
4842 }
4843
4844 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4845 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4846                                         u64 start)
4847 {
4848         struct extent_buffer *eb, *exists = NULL;
4849         int ret;
4850
4851         eb = find_extent_buffer(fs_info, start);
4852         if (eb)
4853                 return eb;
4854         eb = alloc_dummy_extent_buffer(fs_info, start);
4855         if (!eb)
4856                 return NULL;
4857         eb->fs_info = fs_info;
4858 again:
4859         ret = radix_tree_preload(GFP_NOFS);
4860         if (ret)
4861                 goto free_eb;
4862         spin_lock(&fs_info->buffer_lock);
4863         ret = radix_tree_insert(&fs_info->buffer_radix,
4864                                 start >> PAGE_SHIFT, eb);
4865         spin_unlock(&fs_info->buffer_lock);
4866         radix_tree_preload_end();
4867         if (ret == -EEXIST) {
4868                 exists = find_extent_buffer(fs_info, start);
4869                 if (exists)
4870                         goto free_eb;
4871                 else
4872                         goto again;
4873         }
4874         check_buffer_tree_ref(eb);
4875         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4876
4877         /*
4878          * We will free dummy extent buffer's if they come into
4879          * free_extent_buffer with a ref count of 2, but if we are using this we
4880          * want the buffers to stay in memory until we're done with them, so
4881          * bump the ref count again.
4882          */
4883         atomic_inc(&eb->refs);
4884         return eb;
4885 free_eb:
4886         btrfs_release_extent_buffer(eb);
4887         return exists;
4888 }
4889 #endif
4890
4891 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4892                                           u64 start)
4893 {
4894         unsigned long len = fs_info->nodesize;
4895         unsigned long num_pages = num_extent_pages(start, len);
4896         unsigned long i;
4897         unsigned long index = start >> PAGE_SHIFT;
4898         struct extent_buffer *eb;
4899         struct extent_buffer *exists = NULL;
4900         struct page *p;
4901         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4902         int uptodate = 1;
4903         int ret;
4904
4905         if (!IS_ALIGNED(start, fs_info->sectorsize)) {
4906                 btrfs_err(fs_info, "bad tree block start %llu", start);
4907                 return ERR_PTR(-EINVAL);
4908         }
4909
4910         eb = find_extent_buffer(fs_info, start);
4911         if (eb)
4912                 return eb;
4913
4914         eb = __alloc_extent_buffer(fs_info, start, len);
4915         if (!eb)
4916                 return ERR_PTR(-ENOMEM);
4917
4918         for (i = 0; i < num_pages; i++, index++) {
4919                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4920                 if (!p) {
4921                         exists = ERR_PTR(-ENOMEM);
4922                         goto free_eb;
4923                 }
4924
4925                 spin_lock(&mapping->private_lock);
4926                 if (PagePrivate(p)) {
4927                         /*
4928                          * We could have already allocated an eb for this page
4929                          * and attached one so lets see if we can get a ref on
4930                          * the existing eb, and if we can we know it's good and
4931                          * we can just return that one, else we know we can just
4932                          * overwrite page->private.
4933                          */
4934                         exists = (struct extent_buffer *)p->private;
4935                         if (atomic_inc_not_zero(&exists->refs)) {
4936                                 spin_unlock(&mapping->private_lock);
4937                                 unlock_page(p);
4938                                 put_page(p);
4939                                 mark_extent_buffer_accessed(exists, p);
4940                                 goto free_eb;
4941                         }
4942                         exists = NULL;
4943
4944                         /*
4945                          * Do this so attach doesn't complain and we need to
4946                          * drop the ref the old guy had.
4947                          */
4948                         ClearPagePrivate(p);
4949                         WARN_ON(PageDirty(p));
4950                         put_page(p);
4951                 }
4952                 attach_extent_buffer_page(eb, p);
4953                 spin_unlock(&mapping->private_lock);
4954                 WARN_ON(PageDirty(p));
4955                 eb->pages[i] = p;
4956                 if (!PageUptodate(p))
4957                         uptodate = 0;
4958
4959                 /*
4960                  * see below about how we avoid a nasty race with release page
4961                  * and why we unlock later
4962                  */
4963         }
4964         if (uptodate)
4965                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4966 again:
4967         ret = radix_tree_preload(GFP_NOFS);
4968         if (ret) {
4969                 exists = ERR_PTR(ret);
4970                 goto free_eb;
4971         }
4972
4973         spin_lock(&fs_info->buffer_lock);
4974         ret = radix_tree_insert(&fs_info->buffer_radix,
4975                                 start >> PAGE_SHIFT, eb);
4976         spin_unlock(&fs_info->buffer_lock);
4977         radix_tree_preload_end();
4978         if (ret == -EEXIST) {
4979                 exists = find_extent_buffer(fs_info, start);
4980                 if (exists)
4981                         goto free_eb;
4982                 else
4983                         goto again;
4984         }
4985         /* add one reference for the tree */
4986         check_buffer_tree_ref(eb);
4987         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4988
4989         /*
4990          * there is a race where release page may have
4991          * tried to find this extent buffer in the radix
4992          * but failed.  It will tell the VM it is safe to
4993          * reclaim the, and it will clear the page private bit.
4994          * We must make sure to set the page private bit properly
4995          * after the extent buffer is in the radix tree so
4996          * it doesn't get lost
4997          */
4998         SetPageChecked(eb->pages[0]);
4999         for (i = 1; i < num_pages; i++) {
5000                 p = eb->pages[i];
5001                 ClearPageChecked(p);
5002                 unlock_page(p);
5003         }
5004         unlock_page(eb->pages[0]);
5005         return eb;
5006
5007 free_eb:
5008         WARN_ON(!atomic_dec_and_test(&eb->refs));
5009         for (i = 0; i < num_pages; i++) {
5010                 if (eb->pages[i])
5011                         unlock_page(eb->pages[i]);
5012         }
5013
5014         btrfs_release_extent_buffer(eb);
5015         return exists;
5016 }
5017
5018 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5019 {
5020         struct extent_buffer *eb =
5021                         container_of(head, struct extent_buffer, rcu_head);
5022
5023         __free_extent_buffer(eb);
5024 }
5025
5026 /* Expects to have eb->eb_lock already held */
5027 static int release_extent_buffer(struct extent_buffer *eb)
5028 {
5029         WARN_ON(atomic_read(&eb->refs) == 0);
5030         if (atomic_dec_and_test(&eb->refs)) {
5031                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5032                         struct btrfs_fs_info *fs_info = eb->fs_info;
5033
5034                         spin_unlock(&eb->refs_lock);
5035
5036                         spin_lock(&fs_info->buffer_lock);
5037                         radix_tree_delete(&fs_info->buffer_radix,
5038                                           eb->start >> PAGE_SHIFT);
5039                         spin_unlock(&fs_info->buffer_lock);
5040                 } else {
5041                         spin_unlock(&eb->refs_lock);
5042                 }
5043
5044                 /* Should be safe to release our pages at this point */
5045                 btrfs_release_extent_buffer_page(eb);
5046 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5047                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5048                         __free_extent_buffer(eb);
5049                         return 1;
5050                 }
5051 #endif
5052                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5053                 return 1;
5054         }
5055         spin_unlock(&eb->refs_lock);
5056
5057         return 0;
5058 }
5059
5060 void free_extent_buffer(struct extent_buffer *eb)
5061 {
5062         int refs;
5063         int old;
5064         if (!eb)
5065                 return;
5066
5067         while (1) {
5068                 refs = atomic_read(&eb->refs);
5069                 if (refs <= 3)
5070                         break;
5071                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5072                 if (old == refs)
5073                         return;
5074         }
5075
5076         spin_lock(&eb->refs_lock);
5077         if (atomic_read(&eb->refs) == 2 &&
5078             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5079                 atomic_dec(&eb->refs);
5080
5081         if (atomic_read(&eb->refs) == 2 &&
5082             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5083             !extent_buffer_under_io(eb) &&
5084             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5085                 atomic_dec(&eb->refs);
5086
5087         /*
5088          * I know this is terrible, but it's temporary until we stop tracking
5089          * the uptodate bits and such for the extent buffers.
5090          */
5091         release_extent_buffer(eb);
5092 }
5093
5094 void free_extent_buffer_stale(struct extent_buffer *eb)
5095 {
5096         if (!eb)
5097                 return;
5098
5099         spin_lock(&eb->refs_lock);
5100         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5101
5102         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5103             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5104                 atomic_dec(&eb->refs);
5105         release_extent_buffer(eb);
5106 }
5107
5108 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5109 {
5110         unsigned long i;
5111         unsigned long num_pages;
5112         struct page *page;
5113
5114         num_pages = num_extent_pages(eb->start, eb->len);
5115
5116         for (i = 0; i < num_pages; i++) {
5117                 page = eb->pages[i];
5118                 if (!PageDirty(page))
5119                         continue;
5120
5121                 lock_page(page);
5122                 WARN_ON(!PagePrivate(page));
5123
5124                 clear_page_dirty_for_io(page);
5125                 spin_lock_irq(&page->mapping->tree_lock);
5126                 if (!PageDirty(page)) {
5127                         radix_tree_tag_clear(&page->mapping->page_tree,
5128                                                 page_index(page),
5129                                                 PAGECACHE_TAG_DIRTY);
5130                 }
5131                 spin_unlock_irq(&page->mapping->tree_lock);
5132                 ClearPageError(page);
5133                 unlock_page(page);
5134         }
5135         WARN_ON(atomic_read(&eb->refs) == 0);
5136 }
5137
5138 int set_extent_buffer_dirty(struct extent_buffer *eb)
5139 {
5140         unsigned long i;
5141         unsigned long num_pages;
5142         int was_dirty = 0;
5143
5144         check_buffer_tree_ref(eb);
5145
5146         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5147
5148         num_pages = num_extent_pages(eb->start, eb->len);
5149         WARN_ON(atomic_read(&eb->refs) == 0);
5150         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5151
5152         for (i = 0; i < num_pages; i++)
5153                 set_page_dirty(eb->pages[i]);
5154         return was_dirty;
5155 }
5156
5157 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5158 {
5159         unsigned long i;
5160         struct page *page;
5161         unsigned long num_pages;
5162
5163         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5164         num_pages = num_extent_pages(eb->start, eb->len);
5165         for (i = 0; i < num_pages; i++) {
5166                 page = eb->pages[i];
5167                 if (page)
5168                         ClearPageUptodate(page);
5169         }
5170 }
5171
5172 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5173 {
5174         unsigned long i;
5175         struct page *page;
5176         unsigned long num_pages;
5177
5178         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5179         num_pages = num_extent_pages(eb->start, eb->len);
5180         for (i = 0; i < num_pages; i++) {
5181                 page = eb->pages[i];
5182                 SetPageUptodate(page);
5183         }
5184 }
5185
5186 int extent_buffer_uptodate(struct extent_buffer *eb)
5187 {
5188         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5189 }
5190
5191 int read_extent_buffer_pages(struct extent_io_tree *tree,
5192                              struct extent_buffer *eb, int wait,
5193                              get_extent_t *get_extent, int mirror_num)
5194 {
5195         unsigned long i;
5196         struct page *page;
5197         int err;
5198         int ret = 0;
5199         int locked_pages = 0;
5200         int all_uptodate = 1;
5201         unsigned long num_pages;
5202         unsigned long num_reads = 0;
5203         struct bio *bio = NULL;
5204         unsigned long bio_flags = 0;
5205
5206         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5207                 return 0;
5208
5209         num_pages = num_extent_pages(eb->start, eb->len);
5210         for (i = 0; i < num_pages; i++) {
5211                 page = eb->pages[i];
5212                 if (wait == WAIT_NONE) {
5213                         if (!trylock_page(page))
5214                                 goto unlock_exit;
5215                 } else {
5216                         lock_page(page);
5217                 }
5218                 locked_pages++;
5219         }
5220         /*
5221          * We need to firstly lock all pages to make sure that
5222          * the uptodate bit of our pages won't be affected by
5223          * clear_extent_buffer_uptodate().
5224          */
5225         for (i = 0; i < num_pages; i++) {
5226                 page = eb->pages[i];
5227                 if (!PageUptodate(page)) {
5228                         num_reads++;
5229                         all_uptodate = 0;
5230                 }
5231         }
5232
5233         if (all_uptodate) {
5234                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5235                 goto unlock_exit;
5236         }
5237
5238         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5239         eb->read_mirror = 0;
5240         atomic_set(&eb->io_pages, num_reads);
5241         for (i = 0; i < num_pages; i++) {
5242                 page = eb->pages[i];
5243
5244                 if (!PageUptodate(page)) {
5245                         if (ret) {
5246                                 atomic_dec(&eb->io_pages);
5247                                 unlock_page(page);
5248                                 continue;
5249                         }
5250
5251                         ClearPageError(page);
5252                         err = __extent_read_full_page(tree, page,
5253                                                       get_extent, &bio,
5254                                                       mirror_num, &bio_flags,
5255                                                       REQ_META);
5256                         if (err) {
5257                                 ret = err;
5258                                 /*
5259                                  * We use &bio in above __extent_read_full_page,
5260                                  * so we ensure that if it returns error, the
5261                                  * current page fails to add itself to bio and
5262                                  * it's been unlocked.
5263                                  *
5264                                  * We must dec io_pages by ourselves.
5265                                  */
5266                                 atomic_dec(&eb->io_pages);
5267                         }
5268                 } else {
5269                         unlock_page(page);
5270                 }
5271         }
5272
5273         if (bio) {
5274                 err = submit_one_bio(bio, mirror_num, bio_flags);
5275                 if (err)
5276                         return err;
5277         }
5278
5279         if (ret || wait != WAIT_COMPLETE)
5280                 return ret;
5281
5282         for (i = 0; i < num_pages; i++) {
5283                 page = eb->pages[i];
5284                 wait_on_page_locked(page);
5285                 if (!PageUptodate(page))
5286                         ret = -EIO;
5287         }
5288
5289         return ret;
5290
5291 unlock_exit:
5292         while (locked_pages > 0) {
5293                 locked_pages--;
5294                 page = eb->pages[locked_pages];
5295                 unlock_page(page);
5296         }
5297         return ret;
5298 }
5299
5300 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5301                         unsigned long start,
5302                         unsigned long len)
5303 {
5304         size_t cur;
5305         size_t offset;
5306         struct page *page;
5307         char *kaddr;
5308         char *dst = (char *)dstv;
5309         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5310         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5311
5312         WARN_ON(start > eb->len);
5313         WARN_ON(start + len > eb->start + eb->len);
5314
5315         offset = (start_offset + start) & (PAGE_SIZE - 1);
5316
5317         while (len > 0) {
5318                 page = eb->pages[i];
5319
5320                 cur = min(len, (PAGE_SIZE - offset));
5321                 kaddr = page_address(page);
5322                 memcpy(dst, kaddr + offset, cur);
5323
5324                 dst += cur;
5325                 len -= cur;
5326                 offset = 0;
5327                 i++;
5328         }
5329 }
5330
5331 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5332                         unsigned long start,
5333                         unsigned long len)
5334 {
5335         size_t cur;
5336         size_t offset;
5337         struct page *page;
5338         char *kaddr;
5339         char __user *dst = (char __user *)dstv;
5340         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5341         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5342         int ret = 0;
5343
5344         WARN_ON(start > eb->len);
5345         WARN_ON(start + len > eb->start + eb->len);
5346
5347         offset = (start_offset + start) & (PAGE_SIZE - 1);
5348
5349         while (len > 0) {
5350                 page = eb->pages[i];
5351
5352                 cur = min(len, (PAGE_SIZE - offset));
5353                 kaddr = page_address(page);
5354                 if (copy_to_user(dst, kaddr + offset, cur)) {
5355                         ret = -EFAULT;
5356                         break;
5357                 }
5358
5359                 dst += cur;
5360                 len -= cur;
5361                 offset = 0;
5362                 i++;
5363         }
5364
5365         return ret;
5366 }
5367
5368 /*
5369  * return 0 if the item is found within a page.
5370  * return 1 if the item spans two pages.
5371  * return -EINVAL otherwise.
5372  */
5373 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5374                                unsigned long min_len, char **map,
5375                                unsigned long *map_start,
5376                                unsigned long *map_len)
5377 {
5378         size_t offset = start & (PAGE_SIZE - 1);
5379         char *kaddr;
5380         struct page *p;
5381         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5382         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5383         unsigned long end_i = (start_offset + start + min_len - 1) >>
5384                 PAGE_SHIFT;
5385
5386         if (i != end_i)
5387                 return 1;
5388
5389         if (i == 0) {
5390                 offset = start_offset;
5391                 *map_start = 0;
5392         } else {
5393                 offset = 0;
5394                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5395         }
5396
5397         if (start + min_len > eb->len) {
5398                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5399                        eb->start, eb->len, start, min_len);
5400                 return -EINVAL;
5401         }
5402
5403         p = eb->pages[i];
5404         kaddr = page_address(p);
5405         *map = kaddr + offset;
5406         *map_len = PAGE_SIZE - offset;
5407         return 0;
5408 }
5409
5410 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5411                           unsigned long start,
5412                           unsigned long len)
5413 {
5414         size_t cur;
5415         size_t offset;
5416         struct page *page;
5417         char *kaddr;
5418         char *ptr = (char *)ptrv;
5419         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5420         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5421         int ret = 0;
5422
5423         WARN_ON(start > eb->len);
5424         WARN_ON(start + len > eb->start + eb->len);
5425
5426         offset = (start_offset + start) & (PAGE_SIZE - 1);
5427
5428         while (len > 0) {
5429                 page = eb->pages[i];
5430
5431                 cur = min(len, (PAGE_SIZE - offset));
5432
5433                 kaddr = page_address(page);
5434                 ret = memcmp(ptr, kaddr + offset, cur);
5435                 if (ret)
5436                         break;
5437
5438                 ptr += cur;
5439                 len -= cur;
5440                 offset = 0;
5441                 i++;
5442         }
5443         return ret;
5444 }
5445
5446 void write_extent_buffer_chunk_tree_uuid(struct extent_buffer *eb,
5447                 const void *srcv)
5448 {
5449         char *kaddr;
5450
5451         WARN_ON(!PageUptodate(eb->pages[0]));
5452         kaddr = page_address(eb->pages[0]);
5453         memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv,
5454                         BTRFS_FSID_SIZE);
5455 }
5456
5457 void write_extent_buffer_fsid(struct extent_buffer *eb, const void *srcv)
5458 {
5459         char *kaddr;
5460
5461         WARN_ON(!PageUptodate(eb->pages[0]));
5462         kaddr = page_address(eb->pages[0]);
5463         memcpy(kaddr + offsetof(struct btrfs_header, fsid), srcv,
5464                         BTRFS_FSID_SIZE);
5465 }
5466
5467 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5468                          unsigned long start, unsigned long len)
5469 {
5470         size_t cur;
5471         size_t offset;
5472         struct page *page;
5473         char *kaddr;
5474         char *src = (char *)srcv;
5475         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5476         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5477
5478         WARN_ON(start > eb->len);
5479         WARN_ON(start + len > eb->start + eb->len);
5480
5481         offset = (start_offset + start) & (PAGE_SIZE - 1);
5482
5483         while (len > 0) {
5484                 page = eb->pages[i];
5485                 WARN_ON(!PageUptodate(page));
5486
5487                 cur = min(len, PAGE_SIZE - offset);
5488                 kaddr = page_address(page);
5489                 memcpy(kaddr + offset, src, cur);
5490
5491                 src += cur;
5492                 len -= cur;
5493                 offset = 0;
5494                 i++;
5495         }
5496 }
5497
5498 void memzero_extent_buffer(struct extent_buffer *eb, unsigned long start,
5499                 unsigned long len)
5500 {
5501         size_t cur;
5502         size_t offset;
5503         struct page *page;
5504         char *kaddr;
5505         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5506         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5507
5508         WARN_ON(start > eb->len);
5509         WARN_ON(start + len > eb->start + eb->len);
5510
5511         offset = (start_offset + start) & (PAGE_SIZE - 1);
5512
5513         while (len > 0) {
5514                 page = eb->pages[i];
5515                 WARN_ON(!PageUptodate(page));
5516
5517                 cur = min(len, PAGE_SIZE - offset);
5518                 kaddr = page_address(page);
5519                 memset(kaddr + offset, 0, cur);
5520
5521                 len -= cur;
5522                 offset = 0;
5523                 i++;
5524         }
5525 }
5526
5527 void copy_extent_buffer_full(struct extent_buffer *dst,
5528                              struct extent_buffer *src)
5529 {
5530         int i;
5531         unsigned num_pages;
5532
5533         ASSERT(dst->len == src->len);
5534
5535         num_pages = num_extent_pages(dst->start, dst->len);
5536         for (i = 0; i < num_pages; i++)
5537                 copy_page(page_address(dst->pages[i]),
5538                                 page_address(src->pages[i]));
5539 }
5540
5541 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5542                         unsigned long dst_offset, unsigned long src_offset,
5543                         unsigned long len)
5544 {
5545         u64 dst_len = dst->len;
5546         size_t cur;
5547         size_t offset;
5548         struct page *page;
5549         char *kaddr;
5550         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5551         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5552
5553         WARN_ON(src->len != dst_len);
5554
5555         offset = (start_offset + dst_offset) &
5556                 (PAGE_SIZE - 1);
5557
5558         while (len > 0) {
5559                 page = dst->pages[i];
5560                 WARN_ON(!PageUptodate(page));
5561
5562                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5563
5564                 kaddr = page_address(page);
5565                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5566
5567                 src_offset += cur;
5568                 len -= cur;
5569                 offset = 0;
5570                 i++;
5571         }
5572 }
5573
5574 void le_bitmap_set(u8 *map, unsigned int start, int len)
5575 {
5576         u8 *p = map + BIT_BYTE(start);
5577         const unsigned int size = start + len;
5578         int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5579         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);
5580
5581         while (len - bits_to_set >= 0) {
5582                 *p |= mask_to_set;
5583                 len -= bits_to_set;
5584                 bits_to_set = BITS_PER_BYTE;
5585                 mask_to_set = ~0;
5586                 p++;
5587         }
5588         if (len) {
5589                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5590                 *p |= mask_to_set;
5591         }
5592 }
5593
5594 void le_bitmap_clear(u8 *map, unsigned int start, int len)
5595 {
5596         u8 *p = map + BIT_BYTE(start);
5597         const unsigned int size = start + len;
5598         int bits_to_clear = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5599         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(start);
5600
5601         while (len - bits_to_clear >= 0) {
5602                 *p &= ~mask_to_clear;
5603                 len -= bits_to_clear;
5604                 bits_to_clear = BITS_PER_BYTE;
5605                 mask_to_clear = ~0;
5606                 p++;
5607         }
5608         if (len) {
5609                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5610                 *p &= ~mask_to_clear;
5611         }
5612 }
5613
5614 /*
5615  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5616  * given bit number
5617  * @eb: the extent buffer
5618  * @start: offset of the bitmap item in the extent buffer
5619  * @nr: bit number
5620  * @page_index: return index of the page in the extent buffer that contains the
5621  * given bit number
5622  * @page_offset: return offset into the page given by page_index
5623  *
5624  * This helper hides the ugliness of finding the byte in an extent buffer which
5625  * contains a given bit.
5626  */
5627 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5628                                     unsigned long start, unsigned long nr,
5629                                     unsigned long *page_index,
5630                                     size_t *page_offset)
5631 {
5632         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5633         size_t byte_offset = BIT_BYTE(nr);
5634         size_t offset;
5635
5636         /*
5637          * The byte we want is the offset of the extent buffer + the offset of
5638          * the bitmap item in the extent buffer + the offset of the byte in the
5639          * bitmap item.
5640          */
5641         offset = start_offset + start + byte_offset;
5642
5643         *page_index = offset >> PAGE_SHIFT;
5644         *page_offset = offset & (PAGE_SIZE - 1);
5645 }
5646
5647 /**
5648  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5649  * @eb: the extent buffer
5650  * @start: offset of the bitmap item in the extent buffer
5651  * @nr: bit number to test
5652  */
5653 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5654                            unsigned long nr)
5655 {
5656         u8 *kaddr;
5657         struct page *page;
5658         unsigned long i;
5659         size_t offset;
5660
5661         eb_bitmap_offset(eb, start, nr, &i, &offset);
5662         page = eb->pages[i];
5663         WARN_ON(!PageUptodate(page));
5664         kaddr = page_address(page);
5665         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5666 }
5667
5668 /**
5669  * extent_buffer_bitmap_set - set an area of a bitmap
5670  * @eb: the extent buffer
5671  * @start: offset of the bitmap item in the extent buffer
5672  * @pos: bit number of the first bit
5673  * @len: number of bits to set
5674  */
5675 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5676                               unsigned long pos, unsigned long len)
5677 {
5678         u8 *kaddr;
5679         struct page *page;
5680         unsigned long i;
5681         size_t offset;
5682         const unsigned int size = pos + len;
5683         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5684         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5685
5686         eb_bitmap_offset(eb, start, pos, &i, &offset);
5687         page = eb->pages[i];
5688         WARN_ON(!PageUptodate(page));
5689         kaddr = page_address(page);
5690
5691         while (len >= bits_to_set) {
5692                 kaddr[offset] |= mask_to_set;
5693                 len -= bits_to_set;
5694                 bits_to_set = BITS_PER_BYTE;
5695                 mask_to_set = ~0;
5696                 if (++offset >= PAGE_SIZE && len > 0) {
5697                         offset = 0;
5698                         page = eb->pages[++i];
5699                         WARN_ON(!PageUptodate(page));
5700                         kaddr = page_address(page);
5701                 }
5702         }
5703         if (len) {
5704                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5705                 kaddr[offset] |= mask_to_set;
5706         }
5707 }
5708
5709
5710 /**
5711  * extent_buffer_bitmap_clear - clear an area of a bitmap
5712  * @eb: the extent buffer
5713  * @start: offset of the bitmap item in the extent buffer
5714  * @pos: bit number of the first bit
5715  * @len: number of bits to clear
5716  */
5717 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5718                                 unsigned long pos, unsigned long len)
5719 {
5720         u8 *kaddr;
5721         struct page *page;
5722         unsigned long i;
5723         size_t offset;
5724         const unsigned int size = pos + len;
5725         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5726         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5727
5728         eb_bitmap_offset(eb, start, pos, &i, &offset);
5729         page = eb->pages[i];
5730         WARN_ON(!PageUptodate(page));
5731         kaddr = page_address(page);
5732
5733         while (len >= bits_to_clear) {
5734                 kaddr[offset] &= ~mask_to_clear;
5735                 len -= bits_to_clear;
5736                 bits_to_clear = BITS_PER_BYTE;
5737                 mask_to_clear = ~0;
5738                 if (++offset >= PAGE_SIZE && len > 0) {
5739                         offset = 0;
5740                         page = eb->pages[++i];
5741                         WARN_ON(!PageUptodate(page));
5742                         kaddr = page_address(page);
5743                 }
5744         }
5745         if (len) {
5746                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5747                 kaddr[offset] &= ~mask_to_clear;
5748         }
5749 }
5750
5751 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5752 {
5753         unsigned long distance = (src > dst) ? src - dst : dst - src;
5754         return distance < len;
5755 }
5756
5757 static void copy_pages(struct page *dst_page, struct page *src_page,
5758                        unsigned long dst_off, unsigned long src_off,
5759                        unsigned long len)
5760 {
5761         char *dst_kaddr = page_address(dst_page);
5762         char *src_kaddr;
5763         int must_memmove = 0;
5764
5765         if (dst_page != src_page) {
5766                 src_kaddr = page_address(src_page);
5767         } else {
5768                 src_kaddr = dst_kaddr;
5769                 if (areas_overlap(src_off, dst_off, len))
5770                         must_memmove = 1;
5771         }
5772
5773         if (must_memmove)
5774                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5775         else
5776                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5777 }
5778
5779 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5780                            unsigned long src_offset, unsigned long len)
5781 {
5782         struct btrfs_fs_info *fs_info = dst->fs_info;
5783         size_t cur;
5784         size_t dst_off_in_page;
5785         size_t src_off_in_page;
5786         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5787         unsigned long dst_i;
5788         unsigned long src_i;
5789
5790         if (src_offset + len > dst->len) {
5791                 btrfs_err(fs_info,
5792                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5793                          src_offset, len, dst->len);
5794                 BUG_ON(1);
5795         }
5796         if (dst_offset + len > dst->len) {
5797                 btrfs_err(fs_info,
5798                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5799                          dst_offset, len, dst->len);
5800                 BUG_ON(1);
5801         }
5802
5803         while (len > 0) {
5804                 dst_off_in_page = (start_offset + dst_offset) &
5805                         (PAGE_SIZE - 1);
5806                 src_off_in_page = (start_offset + src_offset) &
5807                         (PAGE_SIZE - 1);
5808
5809                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5810                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5811
5812                 cur = min(len, (unsigned long)(PAGE_SIZE -
5813                                                src_off_in_page));
5814                 cur = min_t(unsigned long, cur,
5815                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5816
5817                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5818                            dst_off_in_page, src_off_in_page, cur);
5819
5820                 src_offset += cur;
5821                 dst_offset += cur;
5822                 len -= cur;
5823         }
5824 }
5825
5826 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5827                            unsigned long src_offset, unsigned long len)
5828 {
5829         struct btrfs_fs_info *fs_info = dst->fs_info;
5830         size_t cur;
5831         size_t dst_off_in_page;
5832         size_t src_off_in_page;
5833         unsigned long dst_end = dst_offset + len - 1;
5834         unsigned long src_end = src_offset + len - 1;
5835         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5836         unsigned long dst_i;
5837         unsigned long src_i;
5838
5839         if (src_offset + len > dst->len) {
5840                 btrfs_err(fs_info,
5841                           "memmove bogus src_offset %lu move len %lu len %lu",
5842                           src_offset, len, dst->len);
5843                 BUG_ON(1);
5844         }
5845         if (dst_offset + len > dst->len) {
5846                 btrfs_err(fs_info,
5847                           "memmove bogus dst_offset %lu move len %lu len %lu",
5848                           dst_offset, len, dst->len);
5849                 BUG_ON(1);
5850         }
5851         if (dst_offset < src_offset) {
5852                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5853                 return;
5854         }
5855         while (len > 0) {
5856                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5857                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5858
5859                 dst_off_in_page = (start_offset + dst_end) &
5860                         (PAGE_SIZE - 1);
5861                 src_off_in_page = (start_offset + src_end) &
5862                         (PAGE_SIZE - 1);
5863
5864                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5865                 cur = min(cur, dst_off_in_page + 1);
5866                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5867                            dst_off_in_page - cur + 1,
5868                            src_off_in_page - cur + 1, cur);
5869
5870                 dst_end -= cur;
5871                 src_end -= cur;
5872                 len -= cur;
5873         }
5874 }
5875
5876 int try_release_extent_buffer(struct page *page)
5877 {
5878         struct extent_buffer *eb;
5879
5880         /*
5881          * We need to make sure nobody is attaching this page to an eb right
5882          * now.
5883          */
5884         spin_lock(&page->mapping->private_lock);
5885         if (!PagePrivate(page)) {
5886                 spin_unlock(&page->mapping->private_lock);
5887                 return 1;
5888         }
5889
5890         eb = (struct extent_buffer *)page->private;
5891         BUG_ON(!eb);
5892
5893         /*
5894          * This is a little awful but should be ok, we need to make sure that
5895          * the eb doesn't disappear out from under us while we're looking at
5896          * this page.
5897          */
5898         spin_lock(&eb->refs_lock);
5899         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5900                 spin_unlock(&eb->refs_lock);
5901                 spin_unlock(&page->mapping->private_lock);
5902                 return 0;
5903         }
5904         spin_unlock(&page->mapping->private_lock);
5905
5906         /*
5907          * If tree ref isn't set then we know the ref on this eb is a real ref,
5908          * so just return, this page will likely be freed soon anyway.
5909          */
5910         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5911                 spin_unlock(&eb->refs_lock);
5912                 return 0;
5913         }
5914
5915         return release_extent_buffer(eb);
5916 }