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 && !bio->bi_error)
2588                                 uptodate = 1;
2589                 } else {
2590                         /*
2591                          * The generic bio_readpage_error handles errors the
2592                          * following way: If possible, new read requests are
2593                          * created and submitted and will end up in
2594                          * end_bio_extent_readpage as well (if we're lucky, not
2595                          * in the !uptodate case). In that case it returns 0 and
2596                          * we just go on with the next page in our bio. If it
2597                          * can't handle the error it will return -EIO and we
2598                          * remain responsible for that page.
2599                          */
2600                         ret = bio_readpage_error(bio, offset, page, start, end,
2601                                                  mirror);
2602                         if (ret == 0) {
2603                                 uptodate = !bio->bi_error;
2604                                 offset += len;
2605                                 continue;
2606                         }
2607                 }
2608 readpage_ok:
2609                 if (likely(uptodate)) {
2610                         loff_t i_size = i_size_read(inode);
2611                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2612                         unsigned off;
2613
2614                         /* Zero out the end if this page straddles i_size */
2615                         off = i_size & (PAGE_SIZE-1);
2616                         if (page->index == end_index && off)
2617                                 zero_user_segment(page, off, PAGE_SIZE);
2618                         SetPageUptodate(page);
2619                 } else {
2620                         ClearPageUptodate(page);
2621                         SetPageError(page);
2622                 }
2623                 unlock_page(page);
2624                 offset += len;
2625
2626                 if (unlikely(!uptodate)) {
2627                         if (extent_len) {
2628                                 endio_readpage_release_extent(tree,
2629                                                               extent_start,
2630                                                               extent_len, 1);
2631                                 extent_start = 0;
2632                                 extent_len = 0;
2633                         }
2634                         endio_readpage_release_extent(tree, start,
2635                                                       end - start + 1, 0);
2636                 } else if (!extent_len) {
2637                         extent_start = start;
2638                         extent_len = end + 1 - start;
2639                 } else if (extent_start + extent_len == start) {
2640                         extent_len += end + 1 - start;
2641                 } else {
2642                         endio_readpage_release_extent(tree, extent_start,
2643                                                       extent_len, uptodate);
2644                         extent_start = start;
2645                         extent_len = end + 1 - start;
2646                 }
2647         }
2648
2649         if (extent_len)
2650                 endio_readpage_release_extent(tree, extent_start, extent_len,
2651                                               uptodate);
2652         if (io_bio->end_io)
2653                 io_bio->end_io(io_bio, bio->bi_error);
2654         bio_put(bio);
2655 }
2656
2657 /*
2658  * this allocates from the btrfs_bioset.  We're returning a bio right now
2659  * but you can call btrfs_io_bio for the appropriate container_of magic
2660  */
2661 struct bio *
2662 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2663                 gfp_t gfp_flags)
2664 {
2665         struct btrfs_io_bio *btrfs_bio;
2666         struct bio *bio;
2667
2668         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2669
2670         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2671                 while (!bio && (nr_vecs /= 2)) {
2672                         bio = bio_alloc_bioset(gfp_flags,
2673                                                nr_vecs, btrfs_bioset);
2674                 }
2675         }
2676
2677         if (bio) {
2678                 bio->bi_bdev = bdev;
2679                 bio->bi_iter.bi_sector = first_sector;
2680                 btrfs_bio = btrfs_io_bio(bio);
2681                 btrfs_bio->csum = NULL;
2682                 btrfs_bio->csum_allocated = NULL;
2683                 btrfs_bio->end_io = NULL;
2684         }
2685         return bio;
2686 }
2687
2688 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2689 {
2690         struct btrfs_io_bio *btrfs_bio;
2691         struct bio *new;
2692
2693         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2694         if (new) {
2695                 btrfs_bio = btrfs_io_bio(new);
2696                 btrfs_bio->csum = NULL;
2697                 btrfs_bio->csum_allocated = NULL;
2698                 btrfs_bio->end_io = NULL;
2699         }
2700         return new;
2701 }
2702
2703 /* this also allocates from the btrfs_bioset */
2704 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2705 {
2706         struct btrfs_io_bio *btrfs_bio;
2707         struct bio *bio;
2708
2709         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2710         if (bio) {
2711                 btrfs_bio = btrfs_io_bio(bio);
2712                 btrfs_bio->csum = NULL;
2713                 btrfs_bio->csum_allocated = NULL;
2714                 btrfs_bio->end_io = NULL;
2715         }
2716         return bio;
2717 }
2718
2719
2720 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2721                                        unsigned long bio_flags)
2722 {
2723         int ret = 0;
2724         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2725         struct page *page = bvec->bv_page;
2726         struct extent_io_tree *tree = bio->bi_private;
2727         u64 start;
2728
2729         start = page_offset(page) + bvec->bv_offset;
2730
2731         bio->bi_private = NULL;
2732         bio_get(bio);
2733
2734         if (tree->ops)
2735                 ret = tree->ops->submit_bio_hook(page->mapping->host, bio,
2736                                            mirror_num, bio_flags, start);
2737         else
2738                 btrfsic_submit_bio(bio);
2739
2740         bio_put(bio);
2741         return ret;
2742 }
2743
2744 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2745                      unsigned long offset, size_t size, struct bio *bio,
2746                      unsigned long bio_flags)
2747 {
2748         int ret = 0;
2749         if (tree->ops)
2750                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2751                                                 bio_flags);
2752         return ret;
2753
2754 }
2755
2756 static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree,
2757                               struct writeback_control *wbc,
2758                               struct page *page, sector_t sector,
2759                               size_t size, unsigned long offset,
2760                               struct block_device *bdev,
2761                               struct bio **bio_ret,
2762                               bio_end_io_t end_io_func,
2763                               int mirror_num,
2764                               unsigned long prev_bio_flags,
2765                               unsigned long bio_flags,
2766                               bool force_bio_submit)
2767 {
2768         int ret = 0;
2769         struct bio *bio;
2770         int contig = 0;
2771         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2772         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2773
2774         if (bio_ret && *bio_ret) {
2775                 bio = *bio_ret;
2776                 if (old_compressed)
2777                         contig = bio->bi_iter.bi_sector == sector;
2778                 else
2779                         contig = bio_end_sector(bio) == sector;
2780
2781                 if (prev_bio_flags != bio_flags || !contig ||
2782                     force_bio_submit ||
2783                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2784                     bio_add_page(bio, page, page_size, offset) < page_size) {
2785                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2786                         if (ret < 0) {
2787                                 *bio_ret = NULL;
2788                                 return ret;
2789                         }
2790                         bio = NULL;
2791                 } else {
2792                         if (wbc)
2793                                 wbc_account_io(wbc, page, page_size);
2794                         return 0;
2795                 }
2796         }
2797
2798         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2799                         GFP_NOFS | __GFP_HIGH);
2800         if (!bio)
2801                 return -ENOMEM;
2802
2803         bio_add_page(bio, page, page_size, offset);
2804         bio->bi_end_io = end_io_func;
2805         bio->bi_private = tree;
2806         bio_set_op_attrs(bio, op, op_flags);
2807         if (wbc) {
2808                 wbc_init_bio(wbc, bio);
2809                 wbc_account_io(wbc, page, page_size);
2810         }
2811
2812         if (bio_ret)
2813                 *bio_ret = bio;
2814         else
2815                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2816
2817         return ret;
2818 }
2819
2820 static void attach_extent_buffer_page(struct extent_buffer *eb,
2821                                       struct page *page)
2822 {
2823         if (!PagePrivate(page)) {
2824                 SetPagePrivate(page);
2825                 get_page(page);
2826                 set_page_private(page, (unsigned long)eb);
2827         } else {
2828                 WARN_ON(page->private != (unsigned long)eb);
2829         }
2830 }
2831
2832 void set_page_extent_mapped(struct page *page)
2833 {
2834         if (!PagePrivate(page)) {
2835                 SetPagePrivate(page);
2836                 get_page(page);
2837                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2838         }
2839 }
2840
2841 static struct extent_map *
2842 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2843                  u64 start, u64 len, get_extent_t *get_extent,
2844                  struct extent_map **em_cached)
2845 {
2846         struct extent_map *em;
2847
2848         if (em_cached && *em_cached) {
2849                 em = *em_cached;
2850                 if (extent_map_in_tree(em) && start >= em->start &&
2851                     start < extent_map_end(em)) {
2852                         atomic_inc(&em->refs);
2853                         return em;
2854                 }
2855
2856                 free_extent_map(em);
2857                 *em_cached = NULL;
2858         }
2859
2860         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2861         if (em_cached && !IS_ERR_OR_NULL(em)) {
2862                 BUG_ON(*em_cached);
2863                 atomic_inc(&em->refs);
2864                 *em_cached = em;
2865         }
2866         return em;
2867 }
2868 /*
2869  * basic readpage implementation.  Locked extent state structs are inserted
2870  * into the tree that are removed when the IO is done (by the end_io
2871  * handlers)
2872  * XXX JDM: This needs looking at to ensure proper page locking
2873  * return 0 on success, otherwise return error
2874  */
2875 static int __do_readpage(struct extent_io_tree *tree,
2876                          struct page *page,
2877                          get_extent_t *get_extent,
2878                          struct extent_map **em_cached,
2879                          struct bio **bio, int mirror_num,
2880                          unsigned long *bio_flags, int read_flags,
2881                          u64 *prev_em_start)
2882 {
2883         struct inode *inode = page->mapping->host;
2884         u64 start = page_offset(page);
2885         u64 page_end = start + PAGE_SIZE - 1;
2886         u64 end;
2887         u64 cur = start;
2888         u64 extent_offset;
2889         u64 last_byte = i_size_read(inode);
2890         u64 block_start;
2891         u64 cur_end;
2892         sector_t sector;
2893         struct extent_map *em;
2894         struct block_device *bdev;
2895         int ret = 0;
2896         int nr = 0;
2897         size_t pg_offset = 0;
2898         size_t iosize;
2899         size_t disk_io_size;
2900         size_t blocksize = inode->i_sb->s_blocksize;
2901         unsigned long this_bio_flag = 0;
2902
2903         set_page_extent_mapped(page);
2904
2905         end = page_end;
2906         if (!PageUptodate(page)) {
2907                 if (cleancache_get_page(page) == 0) {
2908                         BUG_ON(blocksize != PAGE_SIZE);
2909                         unlock_extent(tree, start, end);
2910                         goto out;
2911                 }
2912         }
2913
2914         if (page->index == last_byte >> PAGE_SHIFT) {
2915                 char *userpage;
2916                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2917
2918                 if (zero_offset) {
2919                         iosize = PAGE_SIZE - zero_offset;
2920                         userpage = kmap_atomic(page);
2921                         memset(userpage + zero_offset, 0, iosize);
2922                         flush_dcache_page(page);
2923                         kunmap_atomic(userpage);
2924                 }
2925         }
2926         while (cur <= end) {
2927                 bool force_bio_submit = false;
2928
2929                 if (cur >= last_byte) {
2930                         char *userpage;
2931                         struct extent_state *cached = NULL;
2932
2933                         iosize = PAGE_SIZE - pg_offset;
2934                         userpage = kmap_atomic(page);
2935                         memset(userpage + pg_offset, 0, iosize);
2936                         flush_dcache_page(page);
2937                         kunmap_atomic(userpage);
2938                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2939                                             &cached, GFP_NOFS);
2940                         unlock_extent_cached(tree, cur,
2941                                              cur + iosize - 1,
2942                                              &cached, GFP_NOFS);
2943                         break;
2944                 }
2945                 em = __get_extent_map(inode, page, pg_offset, cur,
2946                                       end - cur + 1, get_extent, em_cached);
2947                 if (IS_ERR_OR_NULL(em)) {
2948                         SetPageError(page);
2949                         unlock_extent(tree, cur, end);
2950                         break;
2951                 }
2952                 extent_offset = cur - em->start;
2953                 BUG_ON(extent_map_end(em) <= cur);
2954                 BUG_ON(end < cur);
2955
2956                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2957                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2958                         extent_set_compress_type(&this_bio_flag,
2959                                                  em->compress_type);
2960                 }
2961
2962                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2963                 cur_end = min(extent_map_end(em) - 1, end);
2964                 iosize = ALIGN(iosize, blocksize);
2965                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2966                         disk_io_size = em->block_len;
2967                         sector = em->block_start >> 9;
2968                 } else {
2969                         sector = (em->block_start + extent_offset) >> 9;
2970                         disk_io_size = iosize;
2971                 }
2972                 bdev = em->bdev;
2973                 block_start = em->block_start;
2974                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2975                         block_start = EXTENT_MAP_HOLE;
2976
2977                 /*
2978                  * If we have a file range that points to a compressed extent
2979                  * and it's followed by a consecutive file range that points to
2980                  * to the same compressed extent (possibly with a different
2981                  * offset and/or length, so it either points to the whole extent
2982                  * or only part of it), we must make sure we do not submit a
2983                  * single bio to populate the pages for the 2 ranges because
2984                  * this makes the compressed extent read zero out the pages
2985                  * belonging to the 2nd range. Imagine the following scenario:
2986                  *
2987                  *  File layout
2988                  *  [0 - 8K]                     [8K - 24K]
2989                  *    |                               |
2990                  *    |                               |
2991                  * points to extent X,         points to extent X,
2992                  * offset 4K, length of 8K     offset 0, length 16K
2993                  *
2994                  * [extent X, compressed length = 4K uncompressed length = 16K]
2995                  *
2996                  * If the bio to read the compressed extent covers both ranges,
2997                  * it will decompress extent X into the pages belonging to the
2998                  * first range and then it will stop, zeroing out the remaining
2999                  * pages that belong to the other range that points to extent X.
3000                  * So here we make sure we submit 2 bios, one for the first
3001                  * range and another one for the third range. Both will target
3002                  * the same physical extent from disk, but we can't currently
3003                  * make the compressed bio endio callback populate the pages
3004                  * for both ranges because each compressed bio is tightly
3005                  * coupled with a single extent map, and each range can have
3006                  * an extent map with a different offset value relative to the
3007                  * uncompressed data of our extent and different lengths. This
3008                  * is a corner case so we prioritize correctness over
3009                  * non-optimal behavior (submitting 2 bios for the same extent).
3010                  */
3011                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3012                     prev_em_start && *prev_em_start != (u64)-1 &&
3013                     *prev_em_start != em->orig_start)
3014                         force_bio_submit = true;
3015
3016                 if (prev_em_start)
3017                         *prev_em_start = em->orig_start;
3018
3019                 free_extent_map(em);
3020                 em = NULL;
3021
3022                 /* we've found a hole, just zero and go on */
3023                 if (block_start == EXTENT_MAP_HOLE) {
3024                         char *userpage;
3025                         struct extent_state *cached = NULL;
3026
3027                         userpage = kmap_atomic(page);
3028                         memset(userpage + pg_offset, 0, iosize);
3029                         flush_dcache_page(page);
3030                         kunmap_atomic(userpage);
3031
3032                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3033                                             &cached, GFP_NOFS);
3034                         unlock_extent_cached(tree, cur,
3035                                              cur + iosize - 1,
3036                                              &cached, GFP_NOFS);
3037                         cur = cur + iosize;
3038                         pg_offset += iosize;
3039                         continue;
3040                 }
3041                 /* the get_extent function already copied into the page */
3042                 if (test_range_bit(tree, cur, cur_end,
3043                                    EXTENT_UPTODATE, 1, NULL)) {
3044                         check_page_uptodate(tree, page);
3045                         unlock_extent(tree, cur, cur + iosize - 1);
3046                         cur = cur + iosize;
3047                         pg_offset += iosize;
3048                         continue;
3049                 }
3050                 /* we have an inline extent but it didn't get marked up
3051                  * to date.  Error out
3052                  */
3053                 if (block_start == EXTENT_MAP_INLINE) {
3054                         SetPageError(page);
3055                         unlock_extent(tree, cur, cur + iosize - 1);
3056                         cur = cur + iosize;
3057                         pg_offset += iosize;
3058                         continue;
3059                 }
3060
3061                 ret = submit_extent_page(REQ_OP_READ, read_flags, tree, NULL,
3062                                          page, sector, disk_io_size, pg_offset,
3063                                          bdev, bio,
3064                                          end_bio_extent_readpage, mirror_num,
3065                                          *bio_flags,
3066                                          this_bio_flag,
3067                                          force_bio_submit);
3068                 if (!ret) {
3069                         nr++;
3070                         *bio_flags = this_bio_flag;
3071                 } else {
3072                         SetPageError(page);
3073                         unlock_extent(tree, cur, cur + iosize - 1);
3074                         goto out;
3075                 }
3076                 cur = cur + iosize;
3077                 pg_offset += iosize;
3078         }
3079 out:
3080         if (!nr) {
3081                 if (!PageError(page))
3082                         SetPageUptodate(page);
3083                 unlock_page(page);
3084         }
3085         return ret;
3086 }
3087
3088 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3089                                              struct page *pages[], int nr_pages,
3090                                              u64 start, u64 end,
3091                                              get_extent_t *get_extent,
3092                                              struct extent_map **em_cached,
3093                                              struct bio **bio, int mirror_num,
3094                                              unsigned long *bio_flags,
3095                                              u64 *prev_em_start)
3096 {
3097         struct inode *inode;
3098         struct btrfs_ordered_extent *ordered;
3099         int index;
3100
3101         inode = pages[0]->mapping->host;
3102         while (1) {
3103                 lock_extent(tree, start, end);
3104                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3105                                                      end - start + 1);
3106                 if (!ordered)
3107                         break;
3108                 unlock_extent(tree, start, end);
3109                 btrfs_start_ordered_extent(inode, ordered, 1);
3110                 btrfs_put_ordered_extent(ordered);
3111         }
3112
3113         for (index = 0; index < nr_pages; index++) {
3114                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3115                               mirror_num, bio_flags, 0, prev_em_start);
3116                 put_page(pages[index]);
3117         }
3118 }
3119
3120 static void __extent_readpages(struct extent_io_tree *tree,
3121                                struct page *pages[],
3122                                int nr_pages, get_extent_t *get_extent,
3123                                struct extent_map **em_cached,
3124                                struct bio **bio, int mirror_num,
3125                                unsigned long *bio_flags,
3126                                u64 *prev_em_start)
3127 {
3128         u64 start = 0;
3129         u64 end = 0;
3130         u64 page_start;
3131         int index;
3132         int first_index = 0;
3133
3134         for (index = 0; index < nr_pages; index++) {
3135                 page_start = page_offset(pages[index]);
3136                 if (!end) {
3137                         start = page_start;
3138                         end = start + PAGE_SIZE - 1;
3139                         first_index = index;
3140                 } else if (end + 1 == page_start) {
3141                         end += PAGE_SIZE;
3142                 } else {
3143                         __do_contiguous_readpages(tree, &pages[first_index],
3144                                                   index - first_index, start,
3145                                                   end, get_extent, em_cached,
3146                                                   bio, mirror_num, bio_flags,
3147                                                   prev_em_start);
3148                         start = page_start;
3149                         end = start + PAGE_SIZE - 1;
3150                         first_index = index;
3151                 }
3152         }
3153
3154         if (end)
3155                 __do_contiguous_readpages(tree, &pages[first_index],
3156                                           index - first_index, start,
3157                                           end, get_extent, em_cached, bio,
3158                                           mirror_num, bio_flags,
3159                                           prev_em_start);
3160 }
3161
3162 static int __extent_read_full_page(struct extent_io_tree *tree,
3163                                    struct page *page,
3164                                    get_extent_t *get_extent,
3165                                    struct bio **bio, int mirror_num,
3166                                    unsigned long *bio_flags, int read_flags)
3167 {
3168         struct inode *inode = page->mapping->host;
3169         struct btrfs_ordered_extent *ordered;
3170         u64 start = page_offset(page);
3171         u64 end = start + PAGE_SIZE - 1;
3172         int ret;
3173
3174         while (1) {
3175                 lock_extent(tree, start, end);
3176                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3177                                                 PAGE_SIZE);
3178                 if (!ordered)
3179                         break;
3180                 unlock_extent(tree, start, end);
3181                 btrfs_start_ordered_extent(inode, ordered, 1);
3182                 btrfs_put_ordered_extent(ordered);
3183         }
3184
3185         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3186                             bio_flags, read_flags, NULL);
3187         return ret;
3188 }
3189
3190 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3191                             get_extent_t *get_extent, int mirror_num)
3192 {
3193         struct bio *bio = NULL;
3194         unsigned long bio_flags = 0;
3195         int ret;
3196
3197         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3198                                       &bio_flags, 0);
3199         if (bio)
3200                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3201         return ret;
3202 }
3203
3204 static void update_nr_written(struct writeback_control *wbc,
3205                               unsigned long nr_written)
3206 {
3207         wbc->nr_to_write -= nr_written;
3208 }
3209
3210 /*
3211  * helper for __extent_writepage, doing all of the delayed allocation setup.
3212  *
3213  * This returns 1 if our fill_delalloc function did all the work required
3214  * to write the page (copy into inline extent).  In this case the IO has
3215  * been started and the page is already unlocked.
3216  *
3217  * This returns 0 if all went well (page still locked)
3218  * This returns < 0 if there were errors (page still locked)
3219  */
3220 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3221                               struct page *page, struct writeback_control *wbc,
3222                               struct extent_page_data *epd,
3223                               u64 delalloc_start,
3224                               unsigned long *nr_written)
3225 {
3226         struct extent_io_tree *tree = epd->tree;
3227         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3228         u64 nr_delalloc;
3229         u64 delalloc_to_write = 0;
3230         u64 delalloc_end = 0;
3231         int ret;
3232         int page_started = 0;
3233
3234         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3235                 return 0;
3236
3237         while (delalloc_end < page_end) {
3238                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3239                                                page,
3240                                                &delalloc_start,
3241                                                &delalloc_end,
3242                                                BTRFS_MAX_EXTENT_SIZE);
3243                 if (nr_delalloc == 0) {
3244                         delalloc_start = delalloc_end + 1;
3245                         continue;
3246                 }
3247                 ret = tree->ops->fill_delalloc(inode, page,
3248                                                delalloc_start,
3249                                                delalloc_end,
3250                                                &page_started,
3251                                                nr_written);
3252                 /* File system has been set read-only */
3253                 if (ret) {
3254                         SetPageError(page);
3255                         /* fill_delalloc should be return < 0 for error
3256                          * but just in case, we use > 0 here meaning the
3257                          * IO is started, so we don't want to return > 0
3258                          * unless things are going well.
3259                          */
3260                         ret = ret < 0 ? ret : -EIO;
3261                         goto done;
3262                 }
3263                 /*
3264                  * delalloc_end is already one less than the total length, so
3265                  * we don't subtract one from PAGE_SIZE
3266                  */
3267                 delalloc_to_write += (delalloc_end - delalloc_start +
3268                                       PAGE_SIZE) >> PAGE_SHIFT;
3269                 delalloc_start = delalloc_end + 1;
3270         }
3271         if (wbc->nr_to_write < delalloc_to_write) {
3272                 int thresh = 8192;
3273
3274                 if (delalloc_to_write < thresh * 2)
3275                         thresh = delalloc_to_write;
3276                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3277                                          thresh);
3278         }
3279
3280         /* did the fill delalloc function already unlock and start
3281          * the IO?
3282          */
3283         if (page_started) {
3284                 /*
3285                  * we've unlocked the page, so we can't update
3286                  * the mapping's writeback index, just update
3287                  * nr_to_write.
3288                  */
3289                 wbc->nr_to_write -= *nr_written;
3290                 return 1;
3291         }
3292
3293         ret = 0;
3294
3295 done:
3296         return ret;
3297 }
3298
3299 /*
3300  * helper for __extent_writepage.  This calls the writepage start hooks,
3301  * and does the loop to map the page into extents and bios.
3302  *
3303  * We return 1 if the IO is started and the page is unlocked,
3304  * 0 if all went well (page still locked)
3305  * < 0 if there were errors (page still locked)
3306  */
3307 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3308                                  struct page *page,
3309                                  struct writeback_control *wbc,
3310                                  struct extent_page_data *epd,
3311                                  loff_t i_size,
3312                                  unsigned long nr_written,
3313                                  int write_flags, int *nr_ret)
3314 {
3315         struct extent_io_tree *tree = epd->tree;
3316         u64 start = page_offset(page);
3317         u64 page_end = start + PAGE_SIZE - 1;
3318         u64 end;
3319         u64 cur = start;
3320         u64 extent_offset;
3321         u64 block_start;
3322         u64 iosize;
3323         sector_t sector;
3324         struct extent_map *em;
3325         struct block_device *bdev;
3326         size_t pg_offset = 0;
3327         size_t blocksize;
3328         int ret = 0;
3329         int nr = 0;
3330         bool compressed;
3331
3332         if (tree->ops && tree->ops->writepage_start_hook) {
3333                 ret = tree->ops->writepage_start_hook(page, start,
3334                                                       page_end);
3335                 if (ret) {
3336                         /* Fixup worker will requeue */
3337                         if (ret == -EBUSY)
3338                                 wbc->pages_skipped++;
3339                         else
3340                                 redirty_page_for_writepage(wbc, page);
3341
3342                         update_nr_written(wbc, nr_written);
3343                         unlock_page(page);
3344                         return 1;
3345                 }
3346         }
3347
3348         /*
3349          * we don't want to touch the inode after unlocking the page,
3350          * so we update the mapping writeback index now
3351          */
3352         update_nr_written(wbc, nr_written + 1);
3353
3354         end = page_end;
3355         if (i_size <= start) {
3356                 if (tree->ops && tree->ops->writepage_end_io_hook)
3357                         tree->ops->writepage_end_io_hook(page, start,
3358                                                          page_end, NULL, 1);
3359                 goto done;
3360         }
3361
3362         blocksize = inode->i_sb->s_blocksize;
3363
3364         while (cur <= end) {
3365                 u64 em_end;
3366
3367                 if (cur >= i_size) {
3368                         if (tree->ops && tree->ops->writepage_end_io_hook)
3369                                 tree->ops->writepage_end_io_hook(page, cur,
3370                                                          page_end, NULL, 1);
3371                         break;
3372                 }
3373                 em = epd->get_extent(BTRFS_I(inode), page, pg_offset, cur,
3374                                      end - cur + 1, 1);
3375                 if (IS_ERR_OR_NULL(em)) {
3376                         SetPageError(page);
3377                         ret = PTR_ERR_OR_ZERO(em);
3378                         break;
3379                 }
3380
3381                 extent_offset = cur - em->start;
3382                 em_end = extent_map_end(em);
3383                 BUG_ON(em_end <= cur);
3384                 BUG_ON(end < cur);
3385                 iosize = min(em_end - cur, end - cur + 1);
3386                 iosize = ALIGN(iosize, blocksize);
3387                 sector = (em->block_start + extent_offset) >> 9;
3388                 bdev = em->bdev;
3389                 block_start = em->block_start;
3390                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3391                 free_extent_map(em);
3392                 em = NULL;
3393
3394                 /*
3395                  * compressed and inline extents are written through other
3396                  * paths in the FS
3397                  */
3398                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3399                     block_start == EXTENT_MAP_INLINE) {
3400                         /*
3401                          * end_io notification does not happen here for
3402                          * compressed extents
3403                          */
3404                         if (!compressed && tree->ops &&
3405                             tree->ops->writepage_end_io_hook)
3406                                 tree->ops->writepage_end_io_hook(page, cur,
3407                                                          cur + iosize - 1,
3408                                                          NULL, 1);
3409                         else if (compressed) {
3410                                 /* we don't want to end_page_writeback on
3411                                  * a compressed extent.  this happens
3412                                  * elsewhere
3413                                  */
3414                                 nr++;
3415                         }
3416
3417                         cur += iosize;
3418                         pg_offset += iosize;
3419                         continue;
3420                 }
3421
3422                 set_range_writeback(tree, cur, cur + iosize - 1);
3423                 if (!PageWriteback(page)) {
3424                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3425                                    "page %lu not writeback, cur %llu end %llu",
3426                                page->index, cur, end);
3427                 }
3428
3429                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3430                                          page, sector, iosize, pg_offset,
3431                                          bdev, &epd->bio,
3432                                          end_bio_extent_writepage,
3433                                          0, 0, 0, false);
3434                 if (ret) {
3435                         SetPageError(page);
3436                         if (PageWriteback(page))
3437                                 end_page_writeback(page);
3438                 }
3439
3440                 cur = cur + iosize;
3441                 pg_offset += iosize;
3442                 nr++;
3443         }
3444 done:
3445         *nr_ret = nr;
3446         return ret;
3447 }
3448
3449 /*
3450  * the writepage semantics are similar to regular writepage.  extent
3451  * records are inserted to lock ranges in the tree, and as dirty areas
3452  * are found, they are marked writeback.  Then the lock bits are removed
3453  * and the end_io handler clears the writeback ranges
3454  */
3455 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3456                               void *data)
3457 {
3458         struct inode *inode = page->mapping->host;
3459         struct extent_page_data *epd = data;
3460         u64 start = page_offset(page);
3461         u64 page_end = start + PAGE_SIZE - 1;
3462         int ret;
3463         int nr = 0;
3464         size_t pg_offset = 0;
3465         loff_t i_size = i_size_read(inode);
3466         unsigned long end_index = i_size >> PAGE_SHIFT;
3467         int write_flags = 0;
3468         unsigned long nr_written = 0;
3469
3470         if (wbc->sync_mode == WB_SYNC_ALL)
3471                 write_flags = REQ_SYNC;
3472
3473         trace___extent_writepage(page, inode, wbc);
3474
3475         WARN_ON(!PageLocked(page));
3476
3477         ClearPageError(page);
3478
3479         pg_offset = i_size & (PAGE_SIZE - 1);
3480         if (page->index > end_index ||
3481            (page->index == end_index && !pg_offset)) {
3482                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3483                 unlock_page(page);
3484                 return 0;
3485         }
3486
3487         if (page->index == end_index) {
3488                 char *userpage;
3489
3490                 userpage = kmap_atomic(page);
3491                 memset(userpage + pg_offset, 0,
3492                        PAGE_SIZE - pg_offset);
3493                 kunmap_atomic(userpage);
3494                 flush_dcache_page(page);
3495         }
3496
3497         pg_offset = 0;
3498
3499         set_page_extent_mapped(page);
3500
3501         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3502         if (ret == 1)
3503                 goto done_unlocked;
3504         if (ret)
3505                 goto done;
3506
3507         ret = __extent_writepage_io(inode, page, wbc, epd,
3508                                     i_size, nr_written, write_flags, &nr);
3509         if (ret == 1)
3510                 goto done_unlocked;
3511
3512 done:
3513         if (nr == 0) {
3514                 /* make sure the mapping tag for page dirty gets cleared */
3515                 set_page_writeback(page);
3516                 end_page_writeback(page);
3517         }
3518         if (PageError(page)) {
3519                 ret = ret < 0 ? ret : -EIO;
3520                 end_extent_writepage(page, ret, start, page_end);
3521         }
3522         unlock_page(page);
3523         return ret;
3524
3525 done_unlocked:
3526         return 0;
3527 }
3528
3529 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3530 {
3531         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3532                        TASK_UNINTERRUPTIBLE);
3533 }
3534
3535 static noinline_for_stack int
3536 lock_extent_buffer_for_io(struct extent_buffer *eb,
3537                           struct btrfs_fs_info *fs_info,
3538                           struct extent_page_data *epd)
3539 {
3540         unsigned long i, num_pages;
3541         int flush = 0;
3542         int ret = 0;
3543
3544         if (!btrfs_try_tree_write_lock(eb)) {
3545                 flush = 1;
3546                 flush_write_bio(epd);
3547                 btrfs_tree_lock(eb);
3548         }
3549
3550         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3551                 btrfs_tree_unlock(eb);
3552                 if (!epd->sync_io)
3553                         return 0;
3554                 if (!flush) {
3555                         flush_write_bio(epd);
3556                         flush = 1;
3557                 }
3558                 while (1) {
3559                         wait_on_extent_buffer_writeback(eb);
3560                         btrfs_tree_lock(eb);
3561                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3562                                 break;
3563                         btrfs_tree_unlock(eb);
3564                 }
3565         }
3566
3567         /*
3568          * We need to do this to prevent races in people who check if the eb is
3569          * under IO since we can end up having no IO bits set for a short period
3570          * of time.
3571          */
3572         spin_lock(&eb->refs_lock);
3573         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3574                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3575                 spin_unlock(&eb->refs_lock);
3576                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3577                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3578                                      -eb->len,
3579                                      fs_info->dirty_metadata_batch);
3580                 ret = 1;
3581         } else {
3582                 spin_unlock(&eb->refs_lock);
3583         }
3584
3585         btrfs_tree_unlock(eb);
3586
3587         if (!ret)
3588                 return ret;
3589
3590         num_pages = num_extent_pages(eb->start, eb->len);
3591         for (i = 0; i < num_pages; i++) {
3592                 struct page *p = eb->pages[i];
3593
3594                 if (!trylock_page(p)) {
3595                         if (!flush) {
3596                                 flush_write_bio(epd);
3597                                 flush = 1;
3598                         }
3599                         lock_page(p);
3600                 }
3601         }
3602
3603         return ret;
3604 }
3605
3606 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3607 {
3608         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3609         smp_mb__after_atomic();
3610         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3611 }
3612
3613 static void set_btree_ioerr(struct page *page)
3614 {
3615         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3616
3617         SetPageError(page);
3618         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3619                 return;
3620
3621         /*
3622          * If writeback for a btree extent that doesn't belong to a log tree
3623          * failed, increment the counter transaction->eb_write_errors.
3624          * We do this because while the transaction is running and before it's
3625          * committing (when we call filemap_fdata[write|wait]_range against
3626          * the btree inode), we might have
3627          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3628          * returns an error or an error happens during writeback, when we're
3629          * committing the transaction we wouldn't know about it, since the pages
3630          * can be no longer dirty nor marked anymore for writeback (if a
3631          * subsequent modification to the extent buffer didn't happen before the
3632          * transaction commit), which makes filemap_fdata[write|wait]_range not
3633          * able to find the pages tagged with SetPageError at transaction
3634          * commit time. So if this happens we must abort the transaction,
3635          * otherwise we commit a super block with btree roots that point to
3636          * btree nodes/leafs whose content on disk is invalid - either garbage
3637          * or the content of some node/leaf from a past generation that got
3638          * cowed or deleted and is no longer valid.
3639          *
3640          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3641          * not be enough - we need to distinguish between log tree extents vs
3642          * non-log tree extents, and the next filemap_fdatawait_range() call
3643          * will catch and clear such errors in the mapping - and that call might
3644          * be from a log sync and not from a transaction commit. Also, checking
3645          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3646          * not done and would not be reliable - the eb might have been released
3647          * from memory and reading it back again means that flag would not be
3648          * set (since it's a runtime flag, not persisted on disk).
3649          *
3650          * Using the flags below in the btree inode also makes us achieve the
3651          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3652          * writeback for all dirty pages and before filemap_fdatawait_range()
3653          * is called, the writeback for all dirty pages had already finished
3654          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3655          * filemap_fdatawait_range() would return success, as it could not know
3656          * that writeback errors happened (the pages were no longer tagged for
3657          * writeback).
3658          */
3659         switch (eb->log_index) {
3660         case -1:
3661                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3662                 break;
3663         case 0:
3664                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3665                 break;
3666         case 1:
3667                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3668                 break;
3669         default:
3670                 BUG(); /* unexpected, logic error */
3671         }
3672 }
3673
3674 static void end_bio_extent_buffer_writepage(struct bio *bio)
3675 {
3676         struct bio_vec *bvec;
3677         struct extent_buffer *eb;
3678         int i, done;
3679
3680         bio_for_each_segment_all(bvec, bio, i) {
3681                 struct page *page = bvec->bv_page;
3682
3683                 eb = (struct extent_buffer *)page->private;
3684                 BUG_ON(!eb);
3685                 done = atomic_dec_and_test(&eb->io_pages);
3686
3687                 if (bio->bi_error ||
3688                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3689                         ClearPageUptodate(page);
3690                         set_btree_ioerr(page);
3691                 }
3692
3693                 end_page_writeback(page);
3694
3695                 if (!done)
3696                         continue;
3697
3698                 end_extent_buffer_writeback(eb);
3699         }
3700
3701         bio_put(bio);
3702 }
3703
3704 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3705                         struct btrfs_fs_info *fs_info,
3706                         struct writeback_control *wbc,
3707                         struct extent_page_data *epd)
3708 {
3709         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3710         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3711         u64 offset = eb->start;
3712         u32 nritems;
3713         unsigned long i, num_pages;
3714         unsigned long bio_flags = 0;
3715         unsigned long start, end;
3716         int write_flags = (epd->sync_io ? REQ_SYNC : 0) | REQ_META;
3717         int ret = 0;
3718
3719         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3720         num_pages = num_extent_pages(eb->start, eb->len);
3721         atomic_set(&eb->io_pages, num_pages);
3722         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3723                 bio_flags = EXTENT_BIO_TREE_LOG;
3724
3725         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3726         nritems = btrfs_header_nritems(eb);
3727         if (btrfs_header_level(eb) > 0) {
3728                 end = btrfs_node_key_ptr_offset(nritems);
3729
3730                 memzero_extent_buffer(eb, end, eb->len - end);
3731         } else {
3732                 /*
3733                  * leaf:
3734                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3735                  */
3736                 start = btrfs_item_nr_offset(nritems);
3737                 end = btrfs_leaf_data(eb) + leaf_data_end(fs_info, eb);
3738                 memzero_extent_buffer(eb, start, end - start);
3739         }
3740
3741         for (i = 0; i < num_pages; i++) {
3742                 struct page *p = eb->pages[i];
3743
3744                 clear_page_dirty_for_io(p);
3745                 set_page_writeback(p);
3746                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3747                                          p, offset >> 9, PAGE_SIZE, 0, bdev,
3748                                          &epd->bio,
3749                                          end_bio_extent_buffer_writepage,
3750                                          0, epd->bio_flags, bio_flags, false);
3751                 epd->bio_flags = bio_flags;
3752                 if (ret) {
3753                         set_btree_ioerr(p);
3754                         if (PageWriteback(p))
3755                                 end_page_writeback(p);
3756                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3757                                 end_extent_buffer_writeback(eb);
3758                         ret = -EIO;
3759                         break;
3760                 }
3761                 offset += PAGE_SIZE;
3762                 update_nr_written(wbc, 1);
3763                 unlock_page(p);
3764         }
3765
3766         if (unlikely(ret)) {
3767                 for (; i < num_pages; i++) {
3768                         struct page *p = eb->pages[i];
3769                         clear_page_dirty_for_io(p);
3770                         unlock_page(p);
3771                 }
3772         }
3773
3774         return ret;
3775 }
3776
3777 int btree_write_cache_pages(struct address_space *mapping,
3778                                    struct writeback_control *wbc)
3779 {
3780         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3781         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3782         struct extent_buffer *eb, *prev_eb = NULL;
3783         struct extent_page_data epd = {
3784                 .bio = NULL,
3785                 .tree = tree,
3786                 .extent_locked = 0,
3787                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3788                 .bio_flags = 0,
3789         };
3790         int ret = 0;
3791         int done = 0;
3792         int nr_to_write_done = 0;
3793         struct pagevec pvec;
3794         int nr_pages;
3795         pgoff_t index;
3796         pgoff_t end;            /* Inclusive */
3797         int scanned = 0;
3798         int tag;
3799
3800         pagevec_init(&pvec, 0);
3801         if (wbc->range_cyclic) {
3802                 index = mapping->writeback_index; /* Start from prev offset */
3803                 end = -1;
3804         } else {
3805                 index = wbc->range_start >> PAGE_SHIFT;
3806                 end = wbc->range_end >> PAGE_SHIFT;
3807                 scanned = 1;
3808         }
3809         if (wbc->sync_mode == WB_SYNC_ALL)
3810                 tag = PAGECACHE_TAG_TOWRITE;
3811         else
3812                 tag = PAGECACHE_TAG_DIRTY;
3813 retry:
3814         if (wbc->sync_mode == WB_SYNC_ALL)
3815                 tag_pages_for_writeback(mapping, index, end);
3816         while (!done && !nr_to_write_done && (index <= end) &&
3817                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3818                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3819                 unsigned i;
3820
3821                 scanned = 1;
3822                 for (i = 0; i < nr_pages; i++) {
3823                         struct page *page = pvec.pages[i];
3824
3825                         if (!PagePrivate(page))
3826                                 continue;
3827
3828                         if (!wbc->range_cyclic && page->index > end) {
3829                                 done = 1;
3830                                 break;
3831                         }
3832
3833                         spin_lock(&mapping->private_lock);
3834                         if (!PagePrivate(page)) {
3835                                 spin_unlock(&mapping->private_lock);
3836                                 continue;
3837                         }
3838
3839                         eb = (struct extent_buffer *)page->private;
3840
3841                         /*
3842                          * Shouldn't happen and normally this would be a BUG_ON
3843                          * but no sense in crashing the users box for something
3844                          * we can survive anyway.
3845                          */
3846                         if (WARN_ON(!eb)) {
3847                                 spin_unlock(&mapping->private_lock);
3848                                 continue;
3849                         }
3850
3851                         if (eb == prev_eb) {
3852                                 spin_unlock(&mapping->private_lock);
3853                                 continue;
3854                         }
3855
3856                         ret = atomic_inc_not_zero(&eb->refs);
3857                         spin_unlock(&mapping->private_lock);
3858                         if (!ret)
3859                                 continue;
3860
3861                         prev_eb = eb;
3862                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3863                         if (!ret) {
3864                                 free_extent_buffer(eb);
3865                                 continue;
3866                         }
3867
3868                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3869                         if (ret) {
3870                                 done = 1;
3871                                 free_extent_buffer(eb);
3872                                 break;
3873                         }
3874                         free_extent_buffer(eb);
3875
3876                         /*
3877                          * the filesystem may choose to bump up nr_to_write.
3878                          * We have to make sure to honor the new nr_to_write
3879                          * at any time
3880                          */
3881                         nr_to_write_done = wbc->nr_to_write <= 0;
3882                 }
3883                 pagevec_release(&pvec);
3884                 cond_resched();
3885         }
3886         if (!scanned && !done) {
3887                 /*
3888                  * We hit the last page and there is more work to be done: wrap
3889                  * back to the start of the file
3890                  */
3891                 scanned = 1;
3892                 index = 0;
3893                 goto retry;
3894         }
3895         flush_write_bio(&epd);
3896         return ret;
3897 }
3898
3899 /**
3900  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3901  * @mapping: address space structure to write
3902  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3903  * @writepage: function called for each page
3904  * @data: data passed to writepage function
3905  *
3906  * If a page is already under I/O, write_cache_pages() skips it, even
3907  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3908  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3909  * and msync() need to guarantee that all the data which was dirty at the time
3910  * the call was made get new I/O started against them.  If wbc->sync_mode is
3911  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3912  * existing IO to complete.
3913  */
3914 static int extent_write_cache_pages(struct address_space *mapping,
3915                              struct writeback_control *wbc,
3916                              writepage_t writepage, void *data,
3917                              void (*flush_fn)(void *))
3918 {
3919         struct inode *inode = mapping->host;
3920         int ret = 0;
3921         int done = 0;
3922         int nr_to_write_done = 0;
3923         struct pagevec pvec;
3924         int nr_pages;
3925         pgoff_t index;
3926         pgoff_t end;            /* Inclusive */
3927         pgoff_t done_index;
3928         int range_whole = 0;
3929         int scanned = 0;
3930         int tag;
3931
3932         /*
3933          * We have to hold onto the inode so that ordered extents can do their
3934          * work when the IO finishes.  The alternative to this is failing to add
3935          * an ordered extent if the igrab() fails there and that is a huge pain
3936          * to deal with, so instead just hold onto the inode throughout the
3937          * writepages operation.  If it fails here we are freeing up the inode
3938          * anyway and we'd rather not waste our time writing out stuff that is
3939          * going to be truncated anyway.
3940          */
3941         if (!igrab(inode))
3942                 return 0;
3943
3944         pagevec_init(&pvec, 0);
3945         if (wbc->range_cyclic) {
3946                 index = mapping->writeback_index; /* Start from prev offset */
3947                 end = -1;
3948         } else {
3949                 index = wbc->range_start >> PAGE_SHIFT;
3950                 end = wbc->range_end >> PAGE_SHIFT;
3951                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3952                         range_whole = 1;
3953                 scanned = 1;
3954         }
3955         if (wbc->sync_mode == WB_SYNC_ALL)
3956                 tag = PAGECACHE_TAG_TOWRITE;
3957         else
3958                 tag = PAGECACHE_TAG_DIRTY;
3959 retry:
3960         if (wbc->sync_mode == WB_SYNC_ALL)
3961                 tag_pages_for_writeback(mapping, index, end);
3962         done_index = index;
3963         while (!done && !nr_to_write_done && (index <= end) &&
3964                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3965                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3966                 unsigned i;
3967
3968                 scanned = 1;
3969                 for (i = 0; i < nr_pages; i++) {
3970                         struct page *page = pvec.pages[i];
3971
3972                         done_index = page->index;
3973                         /*
3974                          * At this point we hold neither mapping->tree_lock nor
3975                          * lock on the page itself: the page may be truncated or
3976                          * invalidated (changing page->mapping to NULL), or even
3977                          * swizzled back from swapper_space to tmpfs file
3978                          * mapping
3979                          */
3980                         if (!trylock_page(page)) {
3981                                 flush_fn(data);
3982                                 lock_page(page);
3983                         }
3984
3985                         if (unlikely(page->mapping != mapping)) {
3986                                 unlock_page(page);
3987                                 continue;
3988                         }
3989
3990                         if (!wbc->range_cyclic && page->index > end) {
3991                                 done = 1;
3992                                 unlock_page(page);
3993                                 continue;
3994                         }
3995
3996                         if (wbc->sync_mode != WB_SYNC_NONE) {
3997                                 if (PageWriteback(page))
3998                                         flush_fn(data);
3999                                 wait_on_page_writeback(page);
4000                         }
4001
4002                         if (PageWriteback(page) ||
4003                             !clear_page_dirty_for_io(page)) {
4004                                 unlock_page(page);
4005                                 continue;
4006                         }
4007
4008                         ret = (*writepage)(page, wbc, data);
4009
4010                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4011                                 unlock_page(page);
4012                                 ret = 0;
4013                         }
4014                         if (ret < 0) {
4015                                 /*
4016                                  * done_index is set past this page,
4017                                  * so media errors will not choke
4018                                  * background writeout for the entire
4019                                  * file. This has consequences for
4020                                  * range_cyclic semantics (ie. it may
4021                                  * not be suitable for data integrity
4022                                  * writeout).
4023                                  */
4024                                 done_index = page->index + 1;
4025                                 done = 1;
4026                                 break;
4027                         }
4028
4029                         /*
4030                          * the filesystem may choose to bump up nr_to_write.
4031                          * We have to make sure to honor the new nr_to_write
4032                          * at any time
4033                          */
4034                         nr_to_write_done = wbc->nr_to_write <= 0;
4035                 }
4036                 pagevec_release(&pvec);
4037                 cond_resched();
4038         }
4039         if (!scanned && !done) {
4040                 /*
4041                  * We hit the last page and there is more work to be done: wrap
4042                  * back to the start of the file
4043                  */
4044                 scanned = 1;
4045                 index = 0;
4046                 goto retry;
4047         }
4048
4049         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4050                 mapping->writeback_index = done_index;
4051
4052         btrfs_add_delayed_iput(inode);
4053         return ret;
4054 }
4055
4056 static void flush_epd_write_bio(struct extent_page_data *epd)
4057 {
4058         if (epd->bio) {
4059                 int ret;
4060
4061                 bio_set_op_attrs(epd->bio, REQ_OP_WRITE,
4062                                  epd->sync_io ? REQ_SYNC : 0);
4063
4064                 ret = submit_one_bio(epd->bio, 0, epd->bio_flags);
4065                 BUG_ON(ret < 0); /* -ENOMEM */
4066                 epd->bio = NULL;
4067         }
4068 }
4069
4070 static noinline void flush_write_bio(void *data)
4071 {
4072         struct extent_page_data *epd = data;
4073         flush_epd_write_bio(epd);
4074 }
4075
4076 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4077                           get_extent_t *get_extent,
4078                           struct writeback_control *wbc)
4079 {
4080         int ret;
4081         struct extent_page_data epd = {
4082                 .bio = NULL,
4083                 .tree = tree,
4084                 .get_extent = get_extent,
4085                 .extent_locked = 0,
4086                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4087                 .bio_flags = 0,
4088         };
4089
4090         ret = __extent_writepage(page, wbc, &epd);
4091
4092         flush_epd_write_bio(&epd);
4093         return ret;
4094 }
4095
4096 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4097                               u64 start, u64 end, get_extent_t *get_extent,
4098                               int mode)
4099 {
4100         int ret = 0;
4101         struct address_space *mapping = inode->i_mapping;
4102         struct page *page;
4103         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4104                 PAGE_SHIFT;
4105
4106         struct extent_page_data epd = {
4107                 .bio = NULL,
4108                 .tree = tree,
4109                 .get_extent = get_extent,
4110                 .extent_locked = 1,
4111                 .sync_io = mode == WB_SYNC_ALL,
4112                 .bio_flags = 0,
4113         };
4114         struct writeback_control wbc_writepages = {
4115                 .sync_mode      = mode,
4116                 .nr_to_write    = nr_pages * 2,
4117                 .range_start    = start,
4118                 .range_end      = end + 1,
4119         };
4120
4121         while (start <= end) {
4122                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4123                 if (clear_page_dirty_for_io(page))
4124                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4125                 else {
4126                         if (tree->ops && tree->ops->writepage_end_io_hook)
4127                                 tree->ops->writepage_end_io_hook(page, start,
4128                                                  start + PAGE_SIZE - 1,
4129                                                  NULL, 1);
4130                         unlock_page(page);
4131                 }
4132                 put_page(page);
4133                 start += PAGE_SIZE;
4134         }
4135
4136         flush_epd_write_bio(&epd);
4137         return ret;
4138 }
4139
4140 int extent_writepages(struct extent_io_tree *tree,
4141                       struct address_space *mapping,
4142                       get_extent_t *get_extent,
4143                       struct writeback_control *wbc)
4144 {
4145         int ret = 0;
4146         struct extent_page_data epd = {
4147                 .bio = NULL,
4148                 .tree = tree,
4149                 .get_extent = get_extent,
4150                 .extent_locked = 0,
4151                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4152                 .bio_flags = 0,
4153         };
4154
4155         ret = extent_write_cache_pages(mapping, wbc, __extent_writepage, &epd,
4156                                        flush_write_bio);
4157         flush_epd_write_bio(&epd);
4158         return ret;
4159 }
4160
4161 int extent_readpages(struct extent_io_tree *tree,
4162                      struct address_space *mapping,
4163                      struct list_head *pages, unsigned nr_pages,
4164                      get_extent_t get_extent)
4165 {
4166         struct bio *bio = NULL;
4167         unsigned page_idx;
4168         unsigned long bio_flags = 0;
4169         struct page *pagepool[16];
4170         struct page *page;
4171         struct extent_map *em_cached = NULL;
4172         int nr = 0;
4173         u64 prev_em_start = (u64)-1;
4174
4175         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4176                 page = list_entry(pages->prev, struct page, lru);
4177
4178                 prefetchw(&page->flags);
4179                 list_del(&page->lru);
4180                 if (add_to_page_cache_lru(page, mapping,
4181                                         page->index,
4182                                         readahead_gfp_mask(mapping))) {
4183                         put_page(page);
4184                         continue;
4185                 }
4186
4187                 pagepool[nr++] = page;
4188                 if (nr < ARRAY_SIZE(pagepool))
4189                         continue;
4190                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4191                                    &bio, 0, &bio_flags, &prev_em_start);
4192                 nr = 0;
4193         }
4194         if (nr)
4195                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4196                                    &bio, 0, &bio_flags, &prev_em_start);
4197
4198         if (em_cached)
4199                 free_extent_map(em_cached);
4200
4201         BUG_ON(!list_empty(pages));
4202         if (bio)
4203                 return submit_one_bio(bio, 0, bio_flags);
4204         return 0;
4205 }
4206
4207 /*
4208  * basic invalidatepage code, this waits on any locked or writeback
4209  * ranges corresponding to the page, and then deletes any extent state
4210  * records from the tree
4211  */
4212 int extent_invalidatepage(struct extent_io_tree *tree,
4213                           struct page *page, unsigned long offset)
4214 {
4215         struct extent_state *cached_state = NULL;
4216         u64 start = page_offset(page);
4217         u64 end = start + PAGE_SIZE - 1;
4218         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4219
4220         start += ALIGN(offset, blocksize);
4221         if (start > end)
4222                 return 0;
4223
4224         lock_extent_bits(tree, start, end, &cached_state);
4225         wait_on_page_writeback(page);
4226         clear_extent_bit(tree, start, end,
4227                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4228                          EXTENT_DO_ACCOUNTING,
4229                          1, 1, &cached_state, GFP_NOFS);
4230         return 0;
4231 }
4232
4233 /*
4234  * a helper for releasepage, this tests for areas of the page that
4235  * are locked or under IO and drops the related state bits if it is safe
4236  * to drop the page.
4237  */
4238 static int try_release_extent_state(struct extent_map_tree *map,
4239                                     struct extent_io_tree *tree,
4240                                     struct page *page, gfp_t mask)
4241 {
4242         u64 start = page_offset(page);
4243         u64 end = start + PAGE_SIZE - 1;
4244         int ret = 1;
4245
4246         if (test_range_bit(tree, start, end,
4247                            EXTENT_IOBITS, 0, NULL))
4248                 ret = 0;
4249         else {
4250                 /*
4251                  * at this point we can safely clear everything except the
4252                  * locked bit and the nodatasum bit
4253                  */
4254                 ret = clear_extent_bit(tree, start, end,
4255                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4256                                  0, 0, NULL, mask);
4257
4258                 /* if clear_extent_bit failed for enomem reasons,
4259                  * we can't allow the release to continue.
4260                  */
4261                 if (ret < 0)
4262                         ret = 0;
4263                 else
4264                         ret = 1;
4265         }
4266         return ret;
4267 }
4268
4269 /*
4270  * a helper for releasepage.  As long as there are no locked extents
4271  * in the range corresponding to the page, both state records and extent
4272  * map records are removed
4273  */
4274 int try_release_extent_mapping(struct extent_map_tree *map,
4275                                struct extent_io_tree *tree, struct page *page,
4276                                gfp_t mask)
4277 {
4278         struct extent_map *em;
4279         u64 start = page_offset(page);
4280         u64 end = start + PAGE_SIZE - 1;
4281
4282         if (gfpflags_allow_blocking(mask) &&
4283             page->mapping->host->i_size > SZ_16M) {
4284                 u64 len;
4285                 while (start <= end) {
4286                         len = end - start + 1;
4287                         write_lock(&map->lock);
4288                         em = lookup_extent_mapping(map, start, len);
4289                         if (!em) {
4290                                 write_unlock(&map->lock);
4291                                 break;
4292                         }
4293                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4294                             em->start != start) {
4295                                 write_unlock(&map->lock);
4296                                 free_extent_map(em);
4297                                 break;
4298                         }
4299                         if (!test_range_bit(tree, em->start,
4300                                             extent_map_end(em) - 1,
4301                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4302                                             0, NULL)) {
4303                                 remove_extent_mapping(map, em);
4304                                 /* once for the rb tree */
4305                                 free_extent_map(em);
4306                         }
4307                         start = extent_map_end(em);
4308                         write_unlock(&map->lock);
4309
4310                         /* once for us */
4311                         free_extent_map(em);
4312                 }
4313         }
4314         return try_release_extent_state(map, tree, page, mask);
4315 }
4316
4317 /*
4318  * helper function for fiemap, which doesn't want to see any holes.
4319  * This maps until we find something past 'last'
4320  */
4321 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4322                                                 u64 offset,
4323                                                 u64 last,
4324                                                 get_extent_t *get_extent)
4325 {
4326         u64 sectorsize = btrfs_inode_sectorsize(inode);
4327         struct extent_map *em;
4328         u64 len;
4329
4330         if (offset >= last)
4331                 return NULL;
4332
4333         while (1) {
4334                 len = last - offset;
4335                 if (len == 0)
4336                         break;
4337                 len = ALIGN(len, sectorsize);
4338                 em = get_extent(BTRFS_I(inode), NULL, 0, offset, len, 0);
4339                 if (IS_ERR_OR_NULL(em))
4340                         return em;
4341
4342                 /* if this isn't a hole return it */
4343                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4344                     em->block_start != EXTENT_MAP_HOLE) {
4345                         return em;
4346                 }
4347
4348                 /* this is a hole, advance to the next extent */
4349                 offset = extent_map_end(em);
4350                 free_extent_map(em);
4351                 if (offset >= last)
4352                         break;
4353         }
4354         return NULL;
4355 }
4356
4357 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4358                 __u64 start, __u64 len, get_extent_t *get_extent)
4359 {
4360         int ret = 0;
4361         u64 off = start;
4362         u64 max = start + len;
4363         u32 flags = 0;
4364         u32 found_type;
4365         u64 last;
4366         u64 last_for_get_extent = 0;
4367         u64 disko = 0;
4368         u64 isize = i_size_read(inode);
4369         struct btrfs_key found_key;
4370         struct extent_map *em = NULL;
4371         struct extent_state *cached_state = NULL;
4372         struct btrfs_path *path;
4373         struct btrfs_root *root = BTRFS_I(inode)->root;
4374         int end = 0;
4375         u64 em_start = 0;
4376         u64 em_len = 0;
4377         u64 em_end = 0;
4378
4379         if (len == 0)
4380                 return -EINVAL;
4381
4382         path = btrfs_alloc_path();
4383         if (!path)
4384                 return -ENOMEM;
4385         path->leave_spinning = 1;
4386
4387         start = round_down(start, btrfs_inode_sectorsize(inode));
4388         len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
4389
4390         /*
4391          * lookup the last file extent.  We're not using i_size here
4392          * because there might be preallocation past i_size
4393          */
4394         ret = btrfs_lookup_file_extent(NULL, root, path,
4395                         btrfs_ino(BTRFS_I(inode)), -1, 0);
4396         if (ret < 0) {
4397                 btrfs_free_path(path);
4398                 return ret;
4399         } else {
4400                 WARN_ON(!ret);
4401                 if (ret == 1)
4402                         ret = 0;
4403         }
4404
4405         path->slots[0]--;
4406         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4407         found_type = found_key.type;
4408
4409         /* No extents, but there might be delalloc bits */
4410         if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) ||
4411             found_type != BTRFS_EXTENT_DATA_KEY) {
4412                 /* have to trust i_size as the end */
4413                 last = (u64)-1;
4414                 last_for_get_extent = isize;
4415         } else {
4416                 /*
4417                  * remember the start of the last extent.  There are a
4418                  * bunch of different factors that go into the length of the
4419                  * extent, so its much less complex to remember where it started
4420                  */
4421                 last = found_key.offset;
4422                 last_for_get_extent = last + 1;
4423         }
4424         btrfs_release_path(path);
4425
4426         /*
4427          * we might have some extents allocated but more delalloc past those
4428          * extents.  so, we trust isize unless the start of the last extent is
4429          * beyond isize
4430          */
4431         if (last < isize) {
4432                 last = (u64)-1;
4433                 last_for_get_extent = isize;
4434         }
4435
4436         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4437                          &cached_state);
4438
4439         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4440                                    get_extent);
4441         if (!em)
4442                 goto out;
4443         if (IS_ERR(em)) {
4444                 ret = PTR_ERR(em);
4445                 goto out;
4446         }
4447
4448         while (!end) {
4449                 u64 offset_in_extent = 0;
4450
4451                 /* break if the extent we found is outside the range */
4452                 if (em->start >= max || extent_map_end(em) < off)
4453                         break;
4454
4455                 /*
4456                  * get_extent may return an extent that starts before our
4457                  * requested range.  We have to make sure the ranges
4458                  * we return to fiemap always move forward and don't
4459                  * overlap, so adjust the offsets here
4460                  */
4461                 em_start = max(em->start, off);
4462
4463                 /*
4464                  * record the offset from the start of the extent
4465                  * for adjusting the disk offset below.  Only do this if the
4466                  * extent isn't compressed since our in ram offset may be past
4467                  * what we have actually allocated on disk.
4468                  */
4469                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4470                         offset_in_extent = em_start - em->start;
4471                 em_end = extent_map_end(em);
4472                 em_len = em_end - em_start;
4473                 disko = 0;
4474                 flags = 0;
4475
4476                 /*
4477                  * bump off for our next call to get_extent
4478                  */
4479                 off = extent_map_end(em);
4480                 if (off >= max)
4481                         end = 1;
4482
4483                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4484                         end = 1;
4485                         flags |= FIEMAP_EXTENT_LAST;
4486                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4487                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4488                                   FIEMAP_EXTENT_NOT_ALIGNED);
4489                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4490                         flags |= (FIEMAP_EXTENT_DELALLOC |
4491                                   FIEMAP_EXTENT_UNKNOWN);
4492                 } else if (fieinfo->fi_extents_max) {
4493                         struct btrfs_trans_handle *trans;
4494
4495                         u64 bytenr = em->block_start -
4496                                 (em->start - em->orig_start);
4497
4498                         disko = em->block_start + offset_in_extent;
4499
4500                         /*
4501                          * We need a trans handle to get delayed refs
4502                          */
4503                         trans = btrfs_join_transaction(root);
4504                         /*
4505                          * It's OK if we can't start a trans we can still check
4506                          * from commit_root
4507                          */
4508                         if (IS_ERR(trans))
4509                                 trans = NULL;
4510
4511                         /*
4512                          * As btrfs supports shared space, this information
4513                          * can be exported to userspace tools via
4514                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4515                          * then we're just getting a count and we can skip the
4516                          * lookup stuff.
4517                          */
4518                         ret = btrfs_check_shared(trans, root->fs_info,
4519                                         root->objectid,
4520                                         btrfs_ino(BTRFS_I(inode)), bytenr);
4521                         if (trans)
4522                                 btrfs_end_transaction(trans);
4523                         if (ret < 0)
4524                                 goto out_free;
4525                         if (ret)
4526                                 flags |= FIEMAP_EXTENT_SHARED;
4527                         ret = 0;
4528                 }
4529                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4530                         flags |= FIEMAP_EXTENT_ENCODED;
4531                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4532                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4533
4534                 free_extent_map(em);
4535                 em = NULL;
4536                 if ((em_start >= last) || em_len == (u64)-1 ||
4537                    (last == (u64)-1 && isize <= em_end)) {
4538                         flags |= FIEMAP_EXTENT_LAST;
4539                         end = 1;
4540                 }
4541
4542                 /* now scan forward to see if this is really the last extent. */
4543                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4544                                            get_extent);
4545                 if (IS_ERR(em)) {
4546                         ret = PTR_ERR(em);
4547                         goto out;
4548                 }
4549                 if (!em) {
4550                         flags |= FIEMAP_EXTENT_LAST;
4551                         end = 1;
4552                 }
4553                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4554                                               em_len, flags);
4555                 if (ret) {
4556                         if (ret == 1)
4557                                 ret = 0;
4558                         goto out_free;
4559                 }
4560         }
4561 out_free:
4562         free_extent_map(em);
4563 out:
4564         btrfs_free_path(path);
4565         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4566                              &cached_state, GFP_NOFS);
4567         return ret;
4568 }
4569
4570 static void __free_extent_buffer(struct extent_buffer *eb)
4571 {
4572         btrfs_leak_debug_del(&eb->leak_list);
4573         kmem_cache_free(extent_buffer_cache, eb);
4574 }
4575
4576 int extent_buffer_under_io(struct extent_buffer *eb)
4577 {
4578         return (atomic_read(&eb->io_pages) ||
4579                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4580                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4581 }
4582
4583 /*
4584  * Helper for releasing extent buffer page.
4585  */
4586 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4587 {
4588         unsigned long index;
4589         struct page *page;
4590         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4591
4592         BUG_ON(extent_buffer_under_io(eb));
4593
4594         index = num_extent_pages(eb->start, eb->len);
4595         if (index == 0)
4596                 return;
4597
4598         do {
4599                 index--;
4600                 page = eb->pages[index];
4601                 if (!page)
4602                         continue;
4603                 if (mapped)
4604                         spin_lock(&page->mapping->private_lock);
4605                 /*
4606                  * We do this since we'll remove the pages after we've
4607                  * removed the eb from the radix tree, so we could race
4608                  * and have this page now attached to the new eb.  So
4609                  * only clear page_private if it's still connected to
4610                  * this eb.
4611                  */
4612                 if (PagePrivate(page) &&
4613                     page->private == (unsigned long)eb) {
4614                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4615                         BUG_ON(PageDirty(page));
4616                         BUG_ON(PageWriteback(page));
4617                         /*
4618                          * We need to make sure we haven't be attached
4619                          * to a new eb.
4620                          */
4621                         ClearPagePrivate(page);
4622                         set_page_private(page, 0);
4623                         /* One for the page private */
4624                         put_page(page);
4625                 }
4626
4627                 if (mapped)
4628                         spin_unlock(&page->mapping->private_lock);
4629
4630                 /* One for when we allocated the page */
4631                 put_page(page);
4632         } while (index != 0);
4633 }
4634
4635 /*
4636  * Helper for releasing the extent buffer.
4637  */
4638 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4639 {
4640         btrfs_release_extent_buffer_page(eb);
4641         __free_extent_buffer(eb);
4642 }
4643
4644 static struct extent_buffer *
4645 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4646                       unsigned long len)
4647 {
4648         struct extent_buffer *eb = NULL;
4649
4650         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4651         eb->start = start;
4652         eb->len = len;
4653         eb->fs_info = fs_info;
4654         eb->bflags = 0;
4655         rwlock_init(&eb->lock);
4656         atomic_set(&eb->write_locks, 0);
4657         atomic_set(&eb->read_locks, 0);
4658         atomic_set(&eb->blocking_readers, 0);
4659         atomic_set(&eb->blocking_writers, 0);
4660         atomic_set(&eb->spinning_readers, 0);
4661         atomic_set(&eb->spinning_writers, 0);
4662         eb->lock_nested = 0;
4663         init_waitqueue_head(&eb->write_lock_wq);
4664         init_waitqueue_head(&eb->read_lock_wq);
4665
4666         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4667
4668         spin_lock_init(&eb->refs_lock);
4669         atomic_set(&eb->refs, 1);
4670         atomic_set(&eb->io_pages, 0);
4671
4672         /*
4673          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4674          */
4675         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4676                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4677         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4678
4679         return eb;
4680 }
4681
4682 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4683 {
4684         unsigned long i;
4685         struct page *p;
4686         struct extent_buffer *new;
4687         unsigned long num_pages = num_extent_pages(src->start, src->len);
4688
4689         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4690         if (new == NULL)
4691                 return NULL;
4692
4693         for (i = 0; i < num_pages; i++) {
4694                 p = alloc_page(GFP_NOFS);
4695                 if (!p) {
4696                         btrfs_release_extent_buffer(new);
4697                         return NULL;
4698                 }
4699                 attach_extent_buffer_page(new, p);
4700                 WARN_ON(PageDirty(p));
4701                 SetPageUptodate(p);
4702                 new->pages[i] = p;
4703                 copy_page(page_address(p), page_address(src->pages[i]));
4704         }
4705
4706         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4707         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4708
4709         return new;
4710 }
4711
4712 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4713                                                   u64 start, unsigned long len)
4714 {
4715         struct extent_buffer *eb;
4716         unsigned long num_pages;
4717         unsigned long i;
4718
4719         num_pages = num_extent_pages(start, len);
4720
4721         eb = __alloc_extent_buffer(fs_info, start, len);
4722         if (!eb)
4723                 return NULL;
4724
4725         for (i = 0; i < num_pages; i++) {
4726                 eb->pages[i] = alloc_page(GFP_NOFS);
4727                 if (!eb->pages[i])
4728                         goto err;
4729         }
4730         set_extent_buffer_uptodate(eb);
4731         btrfs_set_header_nritems(eb, 0);
4732         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4733
4734         return eb;
4735 err:
4736         for (; i > 0; i--)
4737                 __free_page(eb->pages[i - 1]);
4738         __free_extent_buffer(eb);
4739         return NULL;
4740 }
4741
4742 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4743                                                 u64 start)
4744 {
4745         return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
4746 }
4747
4748 static void check_buffer_tree_ref(struct extent_buffer *eb)
4749 {
4750         int refs;
4751         /* the ref bit is tricky.  We have to make sure it is set
4752          * if we have the buffer dirty.   Otherwise the
4753          * code to free a buffer can end up dropping a dirty
4754          * page
4755          *
4756          * Once the ref bit is set, it won't go away while the
4757          * buffer is dirty or in writeback, and it also won't
4758          * go away while we have the reference count on the
4759          * eb bumped.
4760          *
4761          * We can't just set the ref bit without bumping the
4762          * ref on the eb because free_extent_buffer might
4763          * see the ref bit and try to clear it.  If this happens
4764          * free_extent_buffer might end up dropping our original
4765          * ref by mistake and freeing the page before we are able
4766          * to add one more ref.
4767          *
4768          * So bump the ref count first, then set the bit.  If someone
4769          * beat us to it, drop the ref we added.
4770          */
4771         refs = atomic_read(&eb->refs);
4772         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4773                 return;
4774
4775         spin_lock(&eb->refs_lock);
4776         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4777                 atomic_inc(&eb->refs);
4778         spin_unlock(&eb->refs_lock);
4779 }
4780
4781 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4782                 struct page *accessed)
4783 {
4784         unsigned long num_pages, i;
4785
4786         check_buffer_tree_ref(eb);
4787
4788         num_pages = num_extent_pages(eb->start, eb->len);
4789         for (i = 0; i < num_pages; i++) {
4790                 struct page *p = eb->pages[i];
4791
4792                 if (p != accessed)
4793                         mark_page_accessed(p);
4794         }
4795 }
4796
4797 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4798                                          u64 start)
4799 {
4800         struct extent_buffer *eb;
4801
4802         rcu_read_lock();
4803         eb = radix_tree_lookup(&fs_info->buffer_radix,
4804                                start >> PAGE_SHIFT);
4805         if (eb && atomic_inc_not_zero(&eb->refs)) {
4806                 rcu_read_unlock();
4807                 /*
4808                  * Lock our eb's refs_lock to avoid races with
4809                  * free_extent_buffer. When we get our eb it might be flagged
4810                  * with EXTENT_BUFFER_STALE and another task running
4811                  * free_extent_buffer might have seen that flag set,
4812                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4813                  * writeback flags not set) and it's still in the tree (flag
4814                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4815                  * of decrementing the extent buffer's reference count twice.
4816                  * So here we could race and increment the eb's reference count,
4817                  * clear its stale flag, mark it as dirty and drop our reference
4818                  * before the other task finishes executing free_extent_buffer,
4819                  * which would later result in an attempt to free an extent
4820                  * buffer that is dirty.
4821                  */
4822                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4823                         spin_lock(&eb->refs_lock);
4824                         spin_unlock(&eb->refs_lock);
4825                 }
4826                 mark_extent_buffer_accessed(eb, NULL);
4827                 return eb;
4828         }
4829         rcu_read_unlock();
4830
4831         return NULL;
4832 }
4833
4834 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4835 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4836                                         u64 start)
4837 {
4838         struct extent_buffer *eb, *exists = NULL;
4839         int ret;
4840
4841         eb = find_extent_buffer(fs_info, start);
4842         if (eb)
4843                 return eb;
4844         eb = alloc_dummy_extent_buffer(fs_info, start);
4845         if (!eb)
4846                 return NULL;
4847         eb->fs_info = fs_info;
4848 again:
4849         ret = radix_tree_preload(GFP_NOFS);
4850         if (ret)
4851                 goto free_eb;
4852         spin_lock(&fs_info->buffer_lock);
4853         ret = radix_tree_insert(&fs_info->buffer_radix,
4854                                 start >> PAGE_SHIFT, eb);
4855         spin_unlock(&fs_info->buffer_lock);
4856         radix_tree_preload_end();
4857         if (ret == -EEXIST) {
4858                 exists = find_extent_buffer(fs_info, start);
4859                 if (exists)
4860                         goto free_eb;
4861                 else
4862                         goto again;
4863         }
4864         check_buffer_tree_ref(eb);
4865         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4866
4867         /*
4868          * We will free dummy extent buffer's if they come into
4869          * free_extent_buffer with a ref count of 2, but if we are using this we
4870          * want the buffers to stay in memory until we're done with them, so
4871          * bump the ref count again.
4872          */
4873         atomic_inc(&eb->refs);
4874         return eb;
4875 free_eb:
4876         btrfs_release_extent_buffer(eb);
4877         return exists;
4878 }
4879 #endif
4880
4881 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4882                                           u64 start)
4883 {
4884         unsigned long len = fs_info->nodesize;
4885         unsigned long num_pages = num_extent_pages(start, len);
4886         unsigned long i;
4887         unsigned long index = start >> PAGE_SHIFT;
4888         struct extent_buffer *eb;
4889         struct extent_buffer *exists = NULL;
4890         struct page *p;
4891         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4892         int uptodate = 1;
4893         int ret;
4894
4895         if (!IS_ALIGNED(start, fs_info->sectorsize)) {
4896                 btrfs_err(fs_info, "bad tree block start %llu", start);
4897                 return ERR_PTR(-EINVAL);
4898         }
4899
4900         eb = find_extent_buffer(fs_info, start);
4901         if (eb)
4902                 return eb;
4903
4904         eb = __alloc_extent_buffer(fs_info, start, len);
4905         if (!eb)
4906                 return ERR_PTR(-ENOMEM);
4907
4908         for (i = 0; i < num_pages; i++, index++) {
4909                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4910                 if (!p) {
4911                         exists = ERR_PTR(-ENOMEM);
4912                         goto free_eb;
4913                 }
4914
4915                 spin_lock(&mapping->private_lock);
4916                 if (PagePrivate(p)) {
4917                         /*
4918                          * We could have already allocated an eb for this page
4919                          * and attached one so lets see if we can get a ref on
4920                          * the existing eb, and if we can we know it's good and
4921                          * we can just return that one, else we know we can just
4922                          * overwrite page->private.
4923                          */
4924                         exists = (struct extent_buffer *)p->private;
4925                         if (atomic_inc_not_zero(&exists->refs)) {
4926                                 spin_unlock(&mapping->private_lock);
4927                                 unlock_page(p);
4928                                 put_page(p);
4929                                 mark_extent_buffer_accessed(exists, p);
4930                                 goto free_eb;
4931                         }
4932                         exists = NULL;
4933
4934                         /*
4935                          * Do this so attach doesn't complain and we need to
4936                          * drop the ref the old guy had.
4937                          */
4938                         ClearPagePrivate(p);
4939                         WARN_ON(PageDirty(p));
4940                         put_page(p);
4941                 }
4942                 attach_extent_buffer_page(eb, p);
4943                 spin_unlock(&mapping->private_lock);
4944                 WARN_ON(PageDirty(p));
4945                 eb->pages[i] = p;
4946                 if (!PageUptodate(p))
4947                         uptodate = 0;
4948
4949                 /*
4950                  * see below about how we avoid a nasty race with release page
4951                  * and why we unlock later
4952                  */
4953         }
4954         if (uptodate)
4955                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4956 again:
4957         ret = radix_tree_preload(GFP_NOFS);
4958         if (ret) {
4959                 exists = ERR_PTR(ret);
4960                 goto free_eb;
4961         }
4962
4963         spin_lock(&fs_info->buffer_lock);
4964         ret = radix_tree_insert(&fs_info->buffer_radix,
4965                                 start >> PAGE_SHIFT, eb);
4966         spin_unlock(&fs_info->buffer_lock);
4967         radix_tree_preload_end();
4968         if (ret == -EEXIST) {
4969                 exists = find_extent_buffer(fs_info, start);
4970                 if (exists)
4971                         goto free_eb;
4972                 else
4973                         goto again;
4974         }
4975         /* add one reference for the tree */
4976         check_buffer_tree_ref(eb);
4977         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4978
4979         /*
4980          * there is a race where release page may have
4981          * tried to find this extent buffer in the radix
4982          * but failed.  It will tell the VM it is safe to
4983          * reclaim the, and it will clear the page private bit.
4984          * We must make sure to set the page private bit properly
4985          * after the extent buffer is in the radix tree so
4986          * it doesn't get lost
4987          */
4988         SetPageChecked(eb->pages[0]);
4989         for (i = 1; i < num_pages; i++) {
4990                 p = eb->pages[i];
4991                 ClearPageChecked(p);
4992                 unlock_page(p);
4993         }
4994         unlock_page(eb->pages[0]);
4995         return eb;
4996
4997 free_eb:
4998         WARN_ON(!atomic_dec_and_test(&eb->refs));
4999         for (i = 0; i < num_pages; i++) {
5000                 if (eb->pages[i])
5001                         unlock_page(eb->pages[i]);
5002         }
5003
5004         btrfs_release_extent_buffer(eb);
5005         return exists;
5006 }
5007
5008 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5009 {
5010         struct extent_buffer *eb =
5011                         container_of(head, struct extent_buffer, rcu_head);
5012
5013         __free_extent_buffer(eb);
5014 }
5015
5016 /* Expects to have eb->eb_lock already held */
5017 static int release_extent_buffer(struct extent_buffer *eb)
5018 {
5019         WARN_ON(atomic_read(&eb->refs) == 0);
5020         if (atomic_dec_and_test(&eb->refs)) {
5021                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5022                         struct btrfs_fs_info *fs_info = eb->fs_info;
5023
5024                         spin_unlock(&eb->refs_lock);
5025
5026                         spin_lock(&fs_info->buffer_lock);
5027                         radix_tree_delete(&fs_info->buffer_radix,
5028                                           eb->start >> PAGE_SHIFT);
5029                         spin_unlock(&fs_info->buffer_lock);
5030                 } else {
5031                         spin_unlock(&eb->refs_lock);
5032                 }
5033
5034                 /* Should be safe to release our pages at this point */
5035                 btrfs_release_extent_buffer_page(eb);
5036 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5037                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5038                         __free_extent_buffer(eb);
5039                         return 1;
5040                 }
5041 #endif
5042                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5043                 return 1;
5044         }
5045         spin_unlock(&eb->refs_lock);
5046
5047         return 0;
5048 }
5049
5050 void free_extent_buffer(struct extent_buffer *eb)
5051 {
5052         int refs;
5053         int old;
5054         if (!eb)
5055                 return;
5056
5057         while (1) {
5058                 refs = atomic_read(&eb->refs);
5059                 if (refs <= 3)
5060                         break;
5061                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5062                 if (old == refs)
5063                         return;
5064         }
5065
5066         spin_lock(&eb->refs_lock);
5067         if (atomic_read(&eb->refs) == 2 &&
5068             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5069                 atomic_dec(&eb->refs);
5070
5071         if (atomic_read(&eb->refs) == 2 &&
5072             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5073             !extent_buffer_under_io(eb) &&
5074             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5075                 atomic_dec(&eb->refs);
5076
5077         /*
5078          * I know this is terrible, but it's temporary until we stop tracking
5079          * the uptodate bits and such for the extent buffers.
5080          */
5081         release_extent_buffer(eb);
5082 }
5083
5084 void free_extent_buffer_stale(struct extent_buffer *eb)
5085 {
5086         if (!eb)
5087                 return;
5088
5089         spin_lock(&eb->refs_lock);
5090         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5091
5092         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5093             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5094                 atomic_dec(&eb->refs);
5095         release_extent_buffer(eb);
5096 }
5097
5098 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5099 {
5100         unsigned long i;
5101         unsigned long num_pages;
5102         struct page *page;
5103
5104         num_pages = num_extent_pages(eb->start, eb->len);
5105
5106         for (i = 0; i < num_pages; i++) {
5107                 page = eb->pages[i];
5108                 if (!PageDirty(page))
5109                         continue;
5110
5111                 lock_page(page);
5112                 WARN_ON(!PagePrivate(page));
5113
5114                 clear_page_dirty_for_io(page);
5115                 spin_lock_irq(&page->mapping->tree_lock);
5116                 if (!PageDirty(page)) {
5117                         radix_tree_tag_clear(&page->mapping->page_tree,
5118                                                 page_index(page),
5119                                                 PAGECACHE_TAG_DIRTY);
5120                 }
5121                 spin_unlock_irq(&page->mapping->tree_lock);
5122                 ClearPageError(page);
5123                 unlock_page(page);
5124         }
5125         WARN_ON(atomic_read(&eb->refs) == 0);
5126 }
5127
5128 int set_extent_buffer_dirty(struct extent_buffer *eb)
5129 {
5130         unsigned long i;
5131         unsigned long num_pages;
5132         int was_dirty = 0;
5133
5134         check_buffer_tree_ref(eb);
5135
5136         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5137
5138         num_pages = num_extent_pages(eb->start, eb->len);
5139         WARN_ON(atomic_read(&eb->refs) == 0);
5140         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5141
5142         for (i = 0; i < num_pages; i++)
5143                 set_page_dirty(eb->pages[i]);
5144         return was_dirty;
5145 }
5146
5147 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5148 {
5149         unsigned long i;
5150         struct page *page;
5151         unsigned long num_pages;
5152
5153         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5154         num_pages = num_extent_pages(eb->start, eb->len);
5155         for (i = 0; i < num_pages; i++) {
5156                 page = eb->pages[i];
5157                 if (page)
5158                         ClearPageUptodate(page);
5159         }
5160 }
5161
5162 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5163 {
5164         unsigned long i;
5165         struct page *page;
5166         unsigned long num_pages;
5167
5168         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5169         num_pages = num_extent_pages(eb->start, eb->len);
5170         for (i = 0; i < num_pages; i++) {
5171                 page = eb->pages[i];
5172                 SetPageUptodate(page);
5173         }
5174 }
5175
5176 int extent_buffer_uptodate(struct extent_buffer *eb)
5177 {
5178         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5179 }
5180
5181 int read_extent_buffer_pages(struct extent_io_tree *tree,
5182                              struct extent_buffer *eb, int wait,
5183                              get_extent_t *get_extent, int mirror_num)
5184 {
5185         unsigned long i;
5186         struct page *page;
5187         int err;
5188         int ret = 0;
5189         int locked_pages = 0;
5190         int all_uptodate = 1;
5191         unsigned long num_pages;
5192         unsigned long num_reads = 0;
5193         struct bio *bio = NULL;
5194         unsigned long bio_flags = 0;
5195
5196         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5197                 return 0;
5198
5199         num_pages = num_extent_pages(eb->start, eb->len);
5200         for (i = 0; i < num_pages; i++) {
5201                 page = eb->pages[i];
5202                 if (wait == WAIT_NONE) {
5203                         if (!trylock_page(page))
5204                                 goto unlock_exit;
5205                 } else {
5206                         lock_page(page);
5207                 }
5208                 locked_pages++;
5209         }
5210         /*
5211          * We need to firstly lock all pages to make sure that
5212          * the uptodate bit of our pages won't be affected by
5213          * clear_extent_buffer_uptodate().
5214          */
5215         for (i = 0; i < num_pages; i++) {
5216                 page = eb->pages[i];
5217                 if (!PageUptodate(page)) {
5218                         num_reads++;
5219                         all_uptodate = 0;
5220                 }
5221         }
5222
5223         if (all_uptodate) {
5224                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5225                 goto unlock_exit;
5226         }
5227
5228         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5229         eb->read_mirror = 0;
5230         atomic_set(&eb->io_pages, num_reads);
5231         for (i = 0; i < num_pages; i++) {
5232                 page = eb->pages[i];
5233
5234                 if (!PageUptodate(page)) {
5235                         if (ret) {
5236                                 atomic_dec(&eb->io_pages);
5237                                 unlock_page(page);
5238                                 continue;
5239                         }
5240
5241                         ClearPageError(page);
5242                         err = __extent_read_full_page(tree, page,
5243                                                       get_extent, &bio,
5244                                                       mirror_num, &bio_flags,
5245                                                       REQ_META);
5246                         if (err) {
5247                                 ret = err;
5248                                 /*
5249                                  * We use &bio in above __extent_read_full_page,
5250                                  * so we ensure that if it returns error, the
5251                                  * current page fails to add itself to bio and
5252                                  * it's been unlocked.
5253                                  *
5254                                  * We must dec io_pages by ourselves.
5255                                  */
5256                                 atomic_dec(&eb->io_pages);
5257                         }
5258                 } else {
5259                         unlock_page(page);
5260                 }
5261         }
5262
5263         if (bio) {
5264                 err = submit_one_bio(bio, mirror_num, bio_flags);
5265                 if (err)
5266                         return err;
5267         }
5268
5269         if (ret || wait != WAIT_COMPLETE)
5270                 return ret;
5271
5272         for (i = 0; i < num_pages; i++) {
5273                 page = eb->pages[i];
5274                 wait_on_page_locked(page);
5275                 if (!PageUptodate(page))
5276                         ret = -EIO;
5277         }
5278
5279         return ret;
5280
5281 unlock_exit:
5282         while (locked_pages > 0) {
5283                 locked_pages--;
5284                 page = eb->pages[locked_pages];
5285                 unlock_page(page);
5286         }
5287         return ret;
5288 }
5289
5290 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5291                         unsigned long start,
5292                         unsigned long len)
5293 {
5294         size_t cur;
5295         size_t offset;
5296         struct page *page;
5297         char *kaddr;
5298         char *dst = (char *)dstv;
5299         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5300         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5301
5302         WARN_ON(start > eb->len);
5303         WARN_ON(start + len > eb->start + eb->len);
5304
5305         offset = (start_offset + start) & (PAGE_SIZE - 1);
5306
5307         while (len > 0) {
5308                 page = eb->pages[i];
5309
5310                 cur = min(len, (PAGE_SIZE - offset));
5311                 kaddr = page_address(page);
5312                 memcpy(dst, kaddr + offset, cur);
5313
5314                 dst += cur;
5315                 len -= cur;
5316                 offset = 0;
5317                 i++;
5318         }
5319 }
5320
5321 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5322                         unsigned long start,
5323                         unsigned long len)
5324 {
5325         size_t cur;
5326         size_t offset;
5327         struct page *page;
5328         char *kaddr;
5329         char __user *dst = (char __user *)dstv;
5330         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5331         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5332         int ret = 0;
5333
5334         WARN_ON(start > eb->len);
5335         WARN_ON(start + len > eb->start + eb->len);
5336
5337         offset = (start_offset + start) & (PAGE_SIZE - 1);
5338
5339         while (len > 0) {
5340                 page = eb->pages[i];
5341
5342                 cur = min(len, (PAGE_SIZE - offset));
5343                 kaddr = page_address(page);
5344                 if (copy_to_user(dst, kaddr + offset, cur)) {
5345                         ret = -EFAULT;
5346                         break;
5347                 }
5348
5349                 dst += cur;
5350                 len -= cur;
5351                 offset = 0;
5352                 i++;
5353         }
5354
5355         return ret;
5356 }
5357
5358 /*
5359  * return 0 if the item is found within a page.
5360  * return 1 if the item spans two pages.
5361  * return -EINVAL otherwise.
5362  */
5363 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5364                                unsigned long min_len, char **map,
5365                                unsigned long *map_start,
5366                                unsigned long *map_len)
5367 {
5368         size_t offset = start & (PAGE_SIZE - 1);
5369         char *kaddr;
5370         struct page *p;
5371         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5372         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5373         unsigned long end_i = (start_offset + start + min_len - 1) >>
5374                 PAGE_SHIFT;
5375
5376         if (i != end_i)
5377                 return 1;
5378
5379         if (i == 0) {
5380                 offset = start_offset;
5381                 *map_start = 0;
5382         } else {
5383                 offset = 0;
5384                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5385         }
5386
5387         if (start + min_len > eb->len) {
5388                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5389                        eb->start, eb->len, start, min_len);
5390                 return -EINVAL;
5391         }
5392
5393         p = eb->pages[i];
5394         kaddr = page_address(p);
5395         *map = kaddr + offset;
5396         *map_len = PAGE_SIZE - offset;
5397         return 0;
5398 }
5399
5400 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5401                           unsigned long start,
5402                           unsigned long len)
5403 {
5404         size_t cur;
5405         size_t offset;
5406         struct page *page;
5407         char *kaddr;
5408         char *ptr = (char *)ptrv;
5409         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5410         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5411         int ret = 0;
5412
5413         WARN_ON(start > eb->len);
5414         WARN_ON(start + len > eb->start + eb->len);
5415
5416         offset = (start_offset + start) & (PAGE_SIZE - 1);
5417
5418         while (len > 0) {
5419                 page = eb->pages[i];
5420
5421                 cur = min(len, (PAGE_SIZE - offset));
5422
5423                 kaddr = page_address(page);
5424                 ret = memcmp(ptr, kaddr + offset, cur);
5425                 if (ret)
5426                         break;
5427
5428                 ptr += cur;
5429                 len -= cur;
5430                 offset = 0;
5431                 i++;
5432         }
5433         return ret;
5434 }
5435
5436 void write_extent_buffer_chunk_tree_uuid(struct extent_buffer *eb,
5437                 const void *srcv)
5438 {
5439         char *kaddr;
5440
5441         WARN_ON(!PageUptodate(eb->pages[0]));
5442         kaddr = page_address(eb->pages[0]);
5443         memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv,
5444                         BTRFS_FSID_SIZE);
5445 }
5446
5447 void write_extent_buffer_fsid(struct extent_buffer *eb, 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, fsid), srcv,
5454                         BTRFS_FSID_SIZE);
5455 }
5456
5457 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5458                          unsigned long start, unsigned long len)
5459 {
5460         size_t cur;
5461         size_t offset;
5462         struct page *page;
5463         char *kaddr;
5464         char *src = (char *)srcv;
5465         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5466         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5467
5468         WARN_ON(start > eb->len);
5469         WARN_ON(start + len > eb->start + eb->len);
5470
5471         offset = (start_offset + start) & (PAGE_SIZE - 1);
5472
5473         while (len > 0) {
5474                 page = eb->pages[i];
5475                 WARN_ON(!PageUptodate(page));
5476
5477                 cur = min(len, PAGE_SIZE - offset);
5478                 kaddr = page_address(page);
5479                 memcpy(kaddr + offset, src, cur);
5480
5481                 src += cur;
5482                 len -= cur;
5483                 offset = 0;
5484                 i++;
5485         }
5486 }
5487
5488 void memzero_extent_buffer(struct extent_buffer *eb, unsigned long start,
5489                 unsigned long len)
5490 {
5491         size_t cur;
5492         size_t offset;
5493         struct page *page;
5494         char *kaddr;
5495         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5496         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5497
5498         WARN_ON(start > eb->len);
5499         WARN_ON(start + len > eb->start + eb->len);
5500
5501         offset = (start_offset + start) & (PAGE_SIZE - 1);
5502
5503         while (len > 0) {
5504                 page = eb->pages[i];
5505                 WARN_ON(!PageUptodate(page));
5506
5507                 cur = min(len, PAGE_SIZE - offset);
5508                 kaddr = page_address(page);
5509                 memset(kaddr + offset, 0, cur);
5510
5511                 len -= cur;
5512                 offset = 0;
5513                 i++;
5514         }
5515 }
5516
5517 void copy_extent_buffer_full(struct extent_buffer *dst,
5518                              struct extent_buffer *src)
5519 {
5520         int i;
5521         unsigned num_pages;
5522
5523         ASSERT(dst->len == src->len);
5524
5525         num_pages = num_extent_pages(dst->start, dst->len);
5526         for (i = 0; i < num_pages; i++)
5527                 copy_page(page_address(dst->pages[i]),
5528                                 page_address(src->pages[i]));
5529 }
5530
5531 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5532                         unsigned long dst_offset, unsigned long src_offset,
5533                         unsigned long len)
5534 {
5535         u64 dst_len = dst->len;
5536         size_t cur;
5537         size_t offset;
5538         struct page *page;
5539         char *kaddr;
5540         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5541         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5542
5543         WARN_ON(src->len != dst_len);
5544
5545         offset = (start_offset + dst_offset) &
5546                 (PAGE_SIZE - 1);
5547
5548         while (len > 0) {
5549                 page = dst->pages[i];
5550                 WARN_ON(!PageUptodate(page));
5551
5552                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5553
5554                 kaddr = page_address(page);
5555                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5556
5557                 src_offset += cur;
5558                 len -= cur;
5559                 offset = 0;
5560                 i++;
5561         }
5562 }
5563
5564 void le_bitmap_set(u8 *map, unsigned int start, int len)
5565 {
5566         u8 *p = map + BIT_BYTE(start);
5567         const unsigned int size = start + len;
5568         int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5569         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);
5570
5571         while (len - bits_to_set >= 0) {
5572                 *p |= mask_to_set;
5573                 len -= bits_to_set;
5574                 bits_to_set = BITS_PER_BYTE;
5575                 mask_to_set = ~0;
5576                 p++;
5577         }
5578         if (len) {
5579                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5580                 *p |= mask_to_set;
5581         }
5582 }
5583
5584 void le_bitmap_clear(u8 *map, unsigned int start, int len)
5585 {
5586         u8 *p = map + BIT_BYTE(start);
5587         const unsigned int size = start + len;
5588         int bits_to_clear = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5589         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(start);
5590
5591         while (len - bits_to_clear >= 0) {
5592                 *p &= ~mask_to_clear;
5593                 len -= bits_to_clear;
5594                 bits_to_clear = BITS_PER_BYTE;
5595                 mask_to_clear = ~0;
5596                 p++;
5597         }
5598         if (len) {
5599                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5600                 *p &= ~mask_to_clear;
5601         }
5602 }
5603
5604 /*
5605  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5606  * given bit number
5607  * @eb: the extent buffer
5608  * @start: offset of the bitmap item in the extent buffer
5609  * @nr: bit number
5610  * @page_index: return index of the page in the extent buffer that contains the
5611  * given bit number
5612  * @page_offset: return offset into the page given by page_index
5613  *
5614  * This helper hides the ugliness of finding the byte in an extent buffer which
5615  * contains a given bit.
5616  */
5617 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5618                                     unsigned long start, unsigned long nr,
5619                                     unsigned long *page_index,
5620                                     size_t *page_offset)
5621 {
5622         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5623         size_t byte_offset = BIT_BYTE(nr);
5624         size_t offset;
5625
5626         /*
5627          * The byte we want is the offset of the extent buffer + the offset of
5628          * the bitmap item in the extent buffer + the offset of the byte in the
5629          * bitmap item.
5630          */
5631         offset = start_offset + start + byte_offset;
5632
5633         *page_index = offset >> PAGE_SHIFT;
5634         *page_offset = offset & (PAGE_SIZE - 1);
5635 }
5636
5637 /**
5638  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5639  * @eb: the extent buffer
5640  * @start: offset of the bitmap item in the extent buffer
5641  * @nr: bit number to test
5642  */
5643 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5644                            unsigned long nr)
5645 {
5646         u8 *kaddr;
5647         struct page *page;
5648         unsigned long i;
5649         size_t offset;
5650
5651         eb_bitmap_offset(eb, start, nr, &i, &offset);
5652         page = eb->pages[i];
5653         WARN_ON(!PageUptodate(page));
5654         kaddr = page_address(page);
5655         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5656 }
5657
5658 /**
5659  * extent_buffer_bitmap_set - set an area of a bitmap
5660  * @eb: the extent buffer
5661  * @start: offset of the bitmap item in the extent buffer
5662  * @pos: bit number of the first bit
5663  * @len: number of bits to set
5664  */
5665 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5666                               unsigned long pos, unsigned long len)
5667 {
5668         u8 *kaddr;
5669         struct page *page;
5670         unsigned long i;
5671         size_t offset;
5672         const unsigned int size = pos + len;
5673         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5674         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5675
5676         eb_bitmap_offset(eb, start, pos, &i, &offset);
5677         page = eb->pages[i];
5678         WARN_ON(!PageUptodate(page));
5679         kaddr = page_address(page);
5680
5681         while (len >= bits_to_set) {
5682                 kaddr[offset] |= mask_to_set;
5683                 len -= bits_to_set;
5684                 bits_to_set = BITS_PER_BYTE;
5685                 mask_to_set = ~0;
5686                 if (++offset >= PAGE_SIZE && len > 0) {
5687                         offset = 0;
5688                         page = eb->pages[++i];
5689                         WARN_ON(!PageUptodate(page));
5690                         kaddr = page_address(page);
5691                 }
5692         }
5693         if (len) {
5694                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5695                 kaddr[offset] |= mask_to_set;
5696         }
5697 }
5698
5699
5700 /**
5701  * extent_buffer_bitmap_clear - clear an area of a bitmap
5702  * @eb: the extent buffer
5703  * @start: offset of the bitmap item in the extent buffer
5704  * @pos: bit number of the first bit
5705  * @len: number of bits to clear
5706  */
5707 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5708                                 unsigned long pos, unsigned long len)
5709 {
5710         u8 *kaddr;
5711         struct page *page;
5712         unsigned long i;
5713         size_t offset;
5714         const unsigned int size = pos + len;
5715         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5716         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5717
5718         eb_bitmap_offset(eb, start, pos, &i, &offset);
5719         page = eb->pages[i];
5720         WARN_ON(!PageUptodate(page));
5721         kaddr = page_address(page);
5722
5723         while (len >= bits_to_clear) {
5724                 kaddr[offset] &= ~mask_to_clear;
5725                 len -= bits_to_clear;
5726                 bits_to_clear = BITS_PER_BYTE;
5727                 mask_to_clear = ~0;
5728                 if (++offset >= PAGE_SIZE && len > 0) {
5729                         offset = 0;
5730                         page = eb->pages[++i];
5731                         WARN_ON(!PageUptodate(page));
5732                         kaddr = page_address(page);
5733                 }
5734         }
5735         if (len) {
5736                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5737                 kaddr[offset] &= ~mask_to_clear;
5738         }
5739 }
5740
5741 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5742 {
5743         unsigned long distance = (src > dst) ? src - dst : dst - src;
5744         return distance < len;
5745 }
5746
5747 static void copy_pages(struct page *dst_page, struct page *src_page,
5748                        unsigned long dst_off, unsigned long src_off,
5749                        unsigned long len)
5750 {
5751         char *dst_kaddr = page_address(dst_page);
5752         char *src_kaddr;
5753         int must_memmove = 0;
5754
5755         if (dst_page != src_page) {
5756                 src_kaddr = page_address(src_page);
5757         } else {
5758                 src_kaddr = dst_kaddr;
5759                 if (areas_overlap(src_off, dst_off, len))
5760                         must_memmove = 1;
5761         }
5762
5763         if (must_memmove)
5764                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5765         else
5766                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5767 }
5768
5769 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5770                            unsigned long src_offset, unsigned long len)
5771 {
5772         struct btrfs_fs_info *fs_info = dst->fs_info;
5773         size_t cur;
5774         size_t dst_off_in_page;
5775         size_t src_off_in_page;
5776         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5777         unsigned long dst_i;
5778         unsigned long src_i;
5779
5780         if (src_offset + len > dst->len) {
5781                 btrfs_err(fs_info,
5782                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5783                          src_offset, len, dst->len);
5784                 BUG_ON(1);
5785         }
5786         if (dst_offset + len > dst->len) {
5787                 btrfs_err(fs_info,
5788                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5789                          dst_offset, len, dst->len);
5790                 BUG_ON(1);
5791         }
5792
5793         while (len > 0) {
5794                 dst_off_in_page = (start_offset + dst_offset) &
5795                         (PAGE_SIZE - 1);
5796                 src_off_in_page = (start_offset + src_offset) &
5797                         (PAGE_SIZE - 1);
5798
5799                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5800                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5801
5802                 cur = min(len, (unsigned long)(PAGE_SIZE -
5803                                                src_off_in_page));
5804                 cur = min_t(unsigned long, cur,
5805                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5806
5807                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5808                            dst_off_in_page, src_off_in_page, cur);
5809
5810                 src_offset += cur;
5811                 dst_offset += cur;
5812                 len -= cur;
5813         }
5814 }
5815
5816 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5817                            unsigned long src_offset, unsigned long len)
5818 {
5819         struct btrfs_fs_info *fs_info = dst->fs_info;
5820         size_t cur;
5821         size_t dst_off_in_page;
5822         size_t src_off_in_page;
5823         unsigned long dst_end = dst_offset + len - 1;
5824         unsigned long src_end = src_offset + len - 1;
5825         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5826         unsigned long dst_i;
5827         unsigned long src_i;
5828
5829         if (src_offset + len > dst->len) {
5830                 btrfs_err(fs_info,
5831                           "memmove bogus src_offset %lu move len %lu len %lu",
5832                           src_offset, len, dst->len);
5833                 BUG_ON(1);
5834         }
5835         if (dst_offset + len > dst->len) {
5836                 btrfs_err(fs_info,
5837                           "memmove bogus dst_offset %lu move len %lu len %lu",
5838                           dst_offset, len, dst->len);
5839                 BUG_ON(1);
5840         }
5841         if (dst_offset < src_offset) {
5842                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5843                 return;
5844         }
5845         while (len > 0) {
5846                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5847                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5848
5849                 dst_off_in_page = (start_offset + dst_end) &
5850                         (PAGE_SIZE - 1);
5851                 src_off_in_page = (start_offset + src_end) &
5852                         (PAGE_SIZE - 1);
5853
5854                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5855                 cur = min(cur, dst_off_in_page + 1);
5856                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5857                            dst_off_in_page - cur + 1,
5858                            src_off_in_page - cur + 1, cur);
5859
5860                 dst_end -= cur;
5861                 src_end -= cur;
5862                 len -= cur;
5863         }
5864 }
5865
5866 int try_release_extent_buffer(struct page *page)
5867 {
5868         struct extent_buffer *eb;
5869
5870         /*
5871          * We need to make sure nobody is attaching this page to an eb right
5872          * now.
5873          */
5874         spin_lock(&page->mapping->private_lock);
5875         if (!PagePrivate(page)) {
5876                 spin_unlock(&page->mapping->private_lock);
5877                 return 1;
5878         }
5879
5880         eb = (struct extent_buffer *)page->private;
5881         BUG_ON(!eb);
5882
5883         /*
5884          * This is a little awful but should be ok, we need to make sure that
5885          * the eb doesn't disappear out from under us while we're looking at
5886          * this page.
5887          */
5888         spin_lock(&eb->refs_lock);
5889         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5890                 spin_unlock(&eb->refs_lock);
5891                 spin_unlock(&page->mapping->private_lock);
5892                 return 0;
5893         }
5894         spin_unlock(&page->mapping->private_lock);
5895
5896         /*
5897          * If tree ref isn't set then we know the ref on this eb is a real ref,
5898          * so just return, this page will likely be freed soon anyway.
5899          */
5900         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5901                 spin_unlock(&eb->refs_lock);
5902                 return 0;
5903         }
5904
5905         return release_extent_buffer(eb);
5906 }