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                        refcount_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         refcount_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 (refcount_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                                 refcount_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                         refcount_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                         refcount_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                         refcount_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         int ret;
2008
2009         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2010         BUG_ON(!mirror_num);
2011
2012         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2013         if (!bio)
2014                 return -EIO;
2015         bio->bi_iter.bi_size = 0;
2016         map_length = length;
2017
2018         /*
2019          * Avoid races with device replace and make sure our bbio has devices
2020          * associated to its stripes that don't go away while we are doing the
2021          * read repair operation.
2022          */
2023         btrfs_bio_counter_inc_blocked(fs_info);
2024         if (btrfs_is_parity_mirror(fs_info, logical, length, mirror_num)) {
2025                 /*
2026                  * Note that we don't use BTRFS_MAP_WRITE because it's supposed
2027                  * to update all raid stripes, but here we just want to correct
2028                  * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
2029                  * stripe's dev and sector.
2030                  */
2031                 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
2032                                       &map_length, &bbio, 0);
2033                 if (ret) {
2034                         btrfs_bio_counter_dec(fs_info);
2035                         bio_put(bio);
2036                         return -EIO;
2037                 }
2038                 ASSERT(bbio->mirror_num == 1);
2039         } else {
2040                 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2041                                       &map_length, &bbio, mirror_num);
2042                 if (ret) {
2043                         btrfs_bio_counter_dec(fs_info);
2044                         bio_put(bio);
2045                         return -EIO;
2046                 }
2047                 BUG_ON(mirror_num != bbio->mirror_num);
2048         }
2049
2050         sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2051         bio->bi_iter.bi_sector = sector;
2052         dev = bbio->stripes[bbio->mirror_num - 1].dev;
2053         btrfs_put_bbio(bbio);
2054         if (!dev || !dev->bdev || !dev->writeable) {
2055                 btrfs_bio_counter_dec(fs_info);
2056                 bio_put(bio);
2057                 return -EIO;
2058         }
2059         bio->bi_bdev = dev->bdev;
2060         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2061         bio_add_page(bio, page, length, pg_offset);
2062
2063         if (btrfsic_submit_bio_wait(bio)) {
2064                 /* try to remap that extent elsewhere? */
2065                 btrfs_bio_counter_dec(fs_info);
2066                 bio_put(bio);
2067                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2068                 return -EIO;
2069         }
2070
2071         btrfs_info_rl_in_rcu(fs_info,
2072                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2073                                   btrfs_ino(inode), start,
2074                                   rcu_str_deref(dev->name), sector);
2075         btrfs_bio_counter_dec(fs_info);
2076         bio_put(bio);
2077         return 0;
2078 }
2079
2080 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2081                          struct extent_buffer *eb, int mirror_num)
2082 {
2083         u64 start = eb->start;
2084         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2085         int ret = 0;
2086
2087         if (fs_info->sb->s_flags & MS_RDONLY)
2088                 return -EROFS;
2089
2090         for (i = 0; i < num_pages; i++) {
2091                 struct page *p = eb->pages[i];
2092
2093                 ret = repair_io_failure(BTRFS_I(fs_info->btree_inode), start,
2094                                         PAGE_SIZE, start, p,
2095                                         start - page_offset(p), mirror_num);
2096                 if (ret)
2097                         break;
2098                 start += PAGE_SIZE;
2099         }
2100
2101         return ret;
2102 }
2103
2104 /*
2105  * each time an IO finishes, we do a fast check in the IO failure tree
2106  * to see if we need to process or clean up an io_failure_record
2107  */
2108 int clean_io_failure(struct btrfs_inode *inode, u64 start, struct page *page,
2109                      unsigned int pg_offset)
2110 {
2111         u64 private;
2112         struct io_failure_record *failrec;
2113         struct btrfs_fs_info *fs_info = inode->root->fs_info;
2114         struct extent_state *state;
2115         int num_copies;
2116         int ret;
2117
2118         private = 0;
2119         ret = count_range_bits(&inode->io_failure_tree, &private,
2120                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2121         if (!ret)
2122                 return 0;
2123
2124         ret = get_state_failrec(&inode->io_failure_tree, start,
2125                         &failrec);
2126         if (ret)
2127                 return 0;
2128
2129         BUG_ON(!failrec->this_mirror);
2130
2131         if (failrec->in_validation) {
2132                 /* there was no real error, just free the record */
2133                 btrfs_debug(fs_info,
2134                         "clean_io_failure: freeing dummy error at %llu",
2135                         failrec->start);
2136                 goto out;
2137         }
2138         if (fs_info->sb->s_flags & MS_RDONLY)
2139                 goto out;
2140
2141         spin_lock(&inode->io_tree.lock);
2142         state = find_first_extent_bit_state(&inode->io_tree,
2143                                             failrec->start,
2144                                             EXTENT_LOCKED);
2145         spin_unlock(&inode->io_tree.lock);
2146
2147         if (state && state->start <= failrec->start &&
2148             state->end >= failrec->start + failrec->len - 1) {
2149                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2150                                               failrec->len);
2151                 if (num_copies > 1)  {
2152                         repair_io_failure(inode, start, failrec->len,
2153                                           failrec->logical, page,
2154                                           pg_offset, failrec->failed_mirror);
2155                 }
2156         }
2157
2158 out:
2159         free_io_failure(inode, failrec);
2160
2161         return 0;
2162 }
2163
2164 /*
2165  * Can be called when
2166  * - hold extent lock
2167  * - under ordered extent
2168  * - the inode is freeing
2169  */
2170 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2171 {
2172         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2173         struct io_failure_record *failrec;
2174         struct extent_state *state, *next;
2175
2176         if (RB_EMPTY_ROOT(&failure_tree->state))
2177                 return;
2178
2179         spin_lock(&failure_tree->lock);
2180         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2181         while (state) {
2182                 if (state->start > end)
2183                         break;
2184
2185                 ASSERT(state->end <= end);
2186
2187                 next = next_state(state);
2188
2189                 failrec = state->failrec;
2190                 free_extent_state(state);
2191                 kfree(failrec);
2192
2193                 state = next;
2194         }
2195         spin_unlock(&failure_tree->lock);
2196 }
2197
2198 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2199                 struct io_failure_record **failrec_ret)
2200 {
2201         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2202         struct io_failure_record *failrec;
2203         struct extent_map *em;
2204         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2205         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2206         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2207         int ret;
2208         u64 logical;
2209
2210         ret = get_state_failrec(failure_tree, start, &failrec);
2211         if (ret) {
2212                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2213                 if (!failrec)
2214                         return -ENOMEM;
2215
2216                 failrec->start = start;
2217                 failrec->len = end - start + 1;
2218                 failrec->this_mirror = 0;
2219                 failrec->bio_flags = 0;
2220                 failrec->in_validation = 0;
2221
2222                 read_lock(&em_tree->lock);
2223                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2224                 if (!em) {
2225                         read_unlock(&em_tree->lock);
2226                         kfree(failrec);
2227                         return -EIO;
2228                 }
2229
2230                 if (em->start > start || em->start + em->len <= start) {
2231                         free_extent_map(em);
2232                         em = NULL;
2233                 }
2234                 read_unlock(&em_tree->lock);
2235                 if (!em) {
2236                         kfree(failrec);
2237                         return -EIO;
2238                 }
2239
2240                 logical = start - em->start;
2241                 logical = em->block_start + logical;
2242                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2243                         logical = em->block_start;
2244                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2245                         extent_set_compress_type(&failrec->bio_flags,
2246                                                  em->compress_type);
2247                 }
2248
2249                 btrfs_debug(fs_info,
2250                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2251                         logical, start, failrec->len);
2252
2253                 failrec->logical = logical;
2254                 free_extent_map(em);
2255
2256                 /* set the bits in the private failure tree */
2257                 ret = set_extent_bits(failure_tree, start, end,
2258                                         EXTENT_LOCKED | EXTENT_DIRTY);
2259                 if (ret >= 0)
2260                         ret = set_state_failrec(failure_tree, start, failrec);
2261                 /* set the bits in the inode's tree */
2262                 if (ret >= 0)
2263                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2264                 if (ret < 0) {
2265                         kfree(failrec);
2266                         return ret;
2267                 }
2268         } else {
2269                 btrfs_debug(fs_info,
2270                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2271                         failrec->logical, failrec->start, failrec->len,
2272                         failrec->in_validation);
2273                 /*
2274                  * when data can be on disk more than twice, add to failrec here
2275                  * (e.g. with a list for failed_mirror) to make
2276                  * clean_io_failure() clean all those errors at once.
2277                  */
2278         }
2279
2280         *failrec_ret = failrec;
2281
2282         return 0;
2283 }
2284
2285 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2286                            struct io_failure_record *failrec, int failed_mirror)
2287 {
2288         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2289         int num_copies;
2290
2291         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2292         if (num_copies == 1) {
2293                 /*
2294                  * we only have a single copy of the data, so don't bother with
2295                  * all the retry and error correction code that follows. no
2296                  * matter what the error is, it is very likely to persist.
2297                  */
2298                 btrfs_debug(fs_info,
2299                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2300                         num_copies, failrec->this_mirror, failed_mirror);
2301                 return 0;
2302         }
2303
2304         /*
2305          * there are two premises:
2306          *      a) deliver good data to the caller
2307          *      b) correct the bad sectors on disk
2308          */
2309         if (failed_bio->bi_vcnt > 1) {
2310                 /*
2311                  * to fulfill b), we need to know the exact failing sectors, as
2312                  * we don't want to rewrite any more than the failed ones. thus,
2313                  * we need separate read requests for the failed bio
2314                  *
2315                  * if the following BUG_ON triggers, our validation request got
2316                  * merged. we need separate requests for our algorithm to work.
2317                  */
2318                 BUG_ON(failrec->in_validation);
2319                 failrec->in_validation = 1;
2320                 failrec->this_mirror = failed_mirror;
2321         } else {
2322                 /*
2323                  * we're ready to fulfill a) and b) alongside. get a good copy
2324                  * of the failed sector and if we succeed, we have setup
2325                  * everything for repair_io_failure to do the rest for us.
2326                  */
2327                 if (failrec->in_validation) {
2328                         BUG_ON(failrec->this_mirror != failed_mirror);
2329                         failrec->in_validation = 0;
2330                         failrec->this_mirror = 0;
2331                 }
2332                 failrec->failed_mirror = failed_mirror;
2333                 failrec->this_mirror++;
2334                 if (failrec->this_mirror == failed_mirror)
2335                         failrec->this_mirror++;
2336         }
2337
2338         if (failrec->this_mirror > num_copies) {
2339                 btrfs_debug(fs_info,
2340                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2341                         num_copies, failrec->this_mirror, failed_mirror);
2342                 return 0;
2343         }
2344
2345         return 1;
2346 }
2347
2348
2349 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2350                                     struct io_failure_record *failrec,
2351                                     struct page *page, int pg_offset, int icsum,
2352                                     bio_end_io_t *endio_func, void *data)
2353 {
2354         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2355         struct bio *bio;
2356         struct btrfs_io_bio *btrfs_failed_bio;
2357         struct btrfs_io_bio *btrfs_bio;
2358
2359         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2360         if (!bio)
2361                 return NULL;
2362
2363         bio->bi_end_io = endio_func;
2364         bio->bi_iter.bi_sector = failrec->logical >> 9;
2365         bio->bi_bdev = fs_info->fs_devices->latest_bdev;
2366         bio->bi_iter.bi_size = 0;
2367         bio->bi_private = data;
2368
2369         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2370         if (btrfs_failed_bio->csum) {
2371                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2372
2373                 btrfs_bio = btrfs_io_bio(bio);
2374                 btrfs_bio->csum = btrfs_bio->csum_inline;
2375                 icsum *= csum_size;
2376                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2377                        csum_size);
2378         }
2379
2380         bio_add_page(bio, page, failrec->len, pg_offset);
2381
2382         return bio;
2383 }
2384
2385 /*
2386  * this is a generic handler for readpage errors (default
2387  * readpage_io_failed_hook). if other copies exist, read those and write back
2388  * good data to the failed position. does not investigate in remapping the
2389  * failed extent elsewhere, hoping the device will be smart enough to do this as
2390  * needed
2391  */
2392
2393 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2394                               struct page *page, u64 start, u64 end,
2395                               int failed_mirror)
2396 {
2397         struct io_failure_record *failrec;
2398         struct inode *inode = page->mapping->host;
2399         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2400         struct bio *bio;
2401         int read_mode = 0;
2402         blk_status_t status;
2403         int ret;
2404
2405         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2406
2407         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2408         if (ret)
2409                 return ret;
2410
2411         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2412         if (!ret) {
2413                 free_io_failure(BTRFS_I(inode), failrec);
2414                 return -EIO;
2415         }
2416
2417         if (failed_bio->bi_vcnt > 1)
2418                 read_mode |= REQ_FAILFAST_DEV;
2419
2420         phy_offset >>= inode->i_sb->s_blocksize_bits;
2421         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2422                                       start - page_offset(page),
2423                                       (int)phy_offset, failed_bio->bi_end_io,
2424                                       NULL);
2425         if (!bio) {
2426                 free_io_failure(BTRFS_I(inode), failrec);
2427                 return -EIO;
2428         }
2429         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2430
2431         btrfs_debug(btrfs_sb(inode->i_sb),
2432                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2433                 read_mode, failrec->this_mirror, failrec->in_validation);
2434
2435         status = tree->ops->submit_bio_hook(inode, bio, failrec->this_mirror,
2436                                          failrec->bio_flags, 0);
2437         if (status) {
2438                 free_io_failure(BTRFS_I(inode), failrec);
2439                 bio_put(bio);
2440                 ret = blk_status_to_errno(status);
2441         }
2442
2443         return ret;
2444 }
2445
2446 /* lots and lots of room for performance fixes in the end_bio funcs */
2447
2448 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2449 {
2450         int uptodate = (err == 0);
2451         struct extent_io_tree *tree;
2452         int ret = 0;
2453
2454         tree = &BTRFS_I(page->mapping->host)->io_tree;
2455
2456         if (tree->ops && tree->ops->writepage_end_io_hook)
2457                 tree->ops->writepage_end_io_hook(page, start, end, NULL,
2458                                 uptodate);
2459
2460         if (!uptodate) {
2461                 ClearPageUptodate(page);
2462                 SetPageError(page);
2463                 ret = err < 0 ? err : -EIO;
2464                 mapping_set_error(page->mapping, ret);
2465         }
2466 }
2467
2468 /*
2469  * after a writepage IO is done, we need to:
2470  * clear the uptodate bits on error
2471  * clear the writeback bits in the extent tree for this IO
2472  * end_page_writeback if the page has no more pending IO
2473  *
2474  * Scheduling is not allowed, so the extent state tree is expected
2475  * to have one and only one object corresponding to this IO.
2476  */
2477 static void end_bio_extent_writepage(struct bio *bio)
2478 {
2479         int error = blk_status_to_errno(bio->bi_status);
2480         struct bio_vec *bvec;
2481         u64 start;
2482         u64 end;
2483         int i;
2484
2485         bio_for_each_segment_all(bvec, bio, i) {
2486                 struct page *page = bvec->bv_page;
2487                 struct inode *inode = page->mapping->host;
2488                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2489
2490                 /* We always issue full-page reads, but if some block
2491                  * in a page fails to read, blk_update_request() will
2492                  * advance bv_offset and adjust bv_len to compensate.
2493                  * Print a warning for nonzero offsets, and an error
2494                  * if they don't add up to a full page.  */
2495                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2496                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2497                                 btrfs_err(fs_info,
2498                                    "partial page write in btrfs with offset %u and length %u",
2499                                         bvec->bv_offset, bvec->bv_len);
2500                         else
2501                                 btrfs_info(fs_info,
2502                                    "incomplete page write in btrfs with offset %u and length %u",
2503                                         bvec->bv_offset, bvec->bv_len);
2504                 }
2505
2506                 start = page_offset(page);
2507                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2508
2509                 end_extent_writepage(page, error, start, end);
2510                 end_page_writeback(page);
2511         }
2512
2513         bio_put(bio);
2514 }
2515
2516 static void
2517 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2518                               int uptodate)
2519 {
2520         struct extent_state *cached = NULL;
2521         u64 end = start + len - 1;
2522
2523         if (uptodate && tree->track_uptodate)
2524                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2525         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2526 }
2527
2528 /*
2529  * after a readpage IO is done, we need to:
2530  * clear the uptodate bits on error
2531  * set the uptodate bits if things worked
2532  * set the page up to date if all extents in the tree are uptodate
2533  * clear the lock bit in the extent tree
2534  * unlock the page if there are no other extents locked for it
2535  *
2536  * Scheduling is not allowed, so the extent state tree is expected
2537  * to have one and only one object corresponding to this IO.
2538  */
2539 static void end_bio_extent_readpage(struct bio *bio)
2540 {
2541         struct bio_vec *bvec;
2542         int uptodate = !bio->bi_status;
2543         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2544         struct extent_io_tree *tree;
2545         u64 offset = 0;
2546         u64 start;
2547         u64 end;
2548         u64 len;
2549         u64 extent_start = 0;
2550         u64 extent_len = 0;
2551         int mirror;
2552         int ret;
2553         int i;
2554
2555         bio_for_each_segment_all(bvec, bio, i) {
2556                 struct page *page = bvec->bv_page;
2557                 struct inode *inode = page->mapping->host;
2558                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2559
2560                 btrfs_debug(fs_info,
2561                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2562                         (u64)bio->bi_iter.bi_sector, bio->bi_status,
2563                         io_bio->mirror_num);
2564                 tree = &BTRFS_I(inode)->io_tree;
2565
2566                 /* We always issue full-page reads, but if some block
2567                  * in a page fails to read, blk_update_request() will
2568                  * advance bv_offset and adjust bv_len to compensate.
2569                  * Print a warning for nonzero offsets, and an error
2570                  * if they don't add up to a full page.  */
2571                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2572                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2573                                 btrfs_err(fs_info,
2574                                         "partial page read in btrfs with offset %u and length %u",
2575                                         bvec->bv_offset, bvec->bv_len);
2576                         else
2577                                 btrfs_info(fs_info,
2578                                         "incomplete page read in btrfs with offset %u and length %u",
2579                                         bvec->bv_offset, bvec->bv_len);
2580                 }
2581
2582                 start = page_offset(page);
2583                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2584                 len = bvec->bv_len;
2585
2586                 mirror = io_bio->mirror_num;
2587                 if (likely(uptodate && tree->ops)) {
2588                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2589                                                               page, start, end,
2590                                                               mirror);
2591                         if (ret)
2592                                 uptodate = 0;
2593                         else
2594                                 clean_io_failure(BTRFS_I(inode), start,
2595                                                 page, 0);
2596                 }
2597
2598                 if (likely(uptodate))
2599                         goto readpage_ok;
2600
2601                 if (tree->ops) {
2602                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2603                         if (ret == -EAGAIN) {
2604                                 /*
2605                                  * Data inode's readpage_io_failed_hook() always
2606                                  * returns -EAGAIN.
2607                                  *
2608                                  * The generic bio_readpage_error handles errors
2609                                  * the following way: If possible, new read
2610                                  * requests are created and submitted and will
2611                                  * end up in end_bio_extent_readpage as well (if
2612                                  * we're lucky, not in the !uptodate case). In
2613                                  * that case it returns 0 and we just go on with
2614                                  * the next page in our bio. If it can't handle
2615                                  * the error it will return -EIO and we remain
2616                                  * responsible for that page.
2617                                  */
2618                                 ret = bio_readpage_error(bio, offset, page,
2619                                                          start, end, mirror);
2620                                 if (ret == 0) {
2621                                         uptodate = !bio->bi_status;
2622                                         offset += len;
2623                                         continue;
2624                                 }
2625                         }
2626
2627                         /*
2628                          * metadata's readpage_io_failed_hook() always returns
2629                          * -EIO and fixes nothing.  -EIO is also returned if
2630                          * data inode error could not be fixed.
2631                          */
2632                         ASSERT(ret == -EIO);
2633                 }
2634 readpage_ok:
2635                 if (likely(uptodate)) {
2636                         loff_t i_size = i_size_read(inode);
2637                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2638                         unsigned off;
2639
2640                         /* Zero out the end if this page straddles i_size */
2641                         off = i_size & (PAGE_SIZE-1);
2642                         if (page->index == end_index && off)
2643                                 zero_user_segment(page, off, PAGE_SIZE);
2644                         SetPageUptodate(page);
2645                 } else {
2646                         ClearPageUptodate(page);
2647                         SetPageError(page);
2648                 }
2649                 unlock_page(page);
2650                 offset += len;
2651
2652                 if (unlikely(!uptodate)) {
2653                         if (extent_len) {
2654                                 endio_readpage_release_extent(tree,
2655                                                               extent_start,
2656                                                               extent_len, 1);
2657                                 extent_start = 0;
2658                                 extent_len = 0;
2659                         }
2660                         endio_readpage_release_extent(tree, start,
2661                                                       end - start + 1, 0);
2662                 } else if (!extent_len) {
2663                         extent_start = start;
2664                         extent_len = end + 1 - start;
2665                 } else if (extent_start + extent_len == start) {
2666                         extent_len += end + 1 - start;
2667                 } else {
2668                         endio_readpage_release_extent(tree, extent_start,
2669                                                       extent_len, uptodate);
2670                         extent_start = start;
2671                         extent_len = end + 1 - start;
2672                 }
2673         }
2674
2675         if (extent_len)
2676                 endio_readpage_release_extent(tree, extent_start, extent_len,
2677                                               uptodate);
2678         if (io_bio->end_io)
2679                 io_bio->end_io(io_bio, blk_status_to_errno(bio->bi_status));
2680         bio_put(bio);
2681 }
2682
2683 /*
2684  * this allocates from the btrfs_bioset.  We're returning a bio right now
2685  * but you can call btrfs_io_bio for the appropriate container_of magic
2686  */
2687 struct bio *
2688 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2689                 gfp_t gfp_flags)
2690 {
2691         struct btrfs_io_bio *btrfs_bio;
2692         struct bio *bio;
2693
2694         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2695
2696         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2697                 while (!bio && (nr_vecs /= 2)) {
2698                         bio = bio_alloc_bioset(gfp_flags,
2699                                                nr_vecs, btrfs_bioset);
2700                 }
2701         }
2702
2703         if (bio) {
2704                 bio->bi_bdev = bdev;
2705                 bio->bi_iter.bi_sector = first_sector;
2706                 btrfs_bio = btrfs_io_bio(bio);
2707                 btrfs_bio->csum = NULL;
2708                 btrfs_bio->csum_allocated = NULL;
2709                 btrfs_bio->end_io = NULL;
2710         }
2711         return bio;
2712 }
2713
2714 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2715 {
2716         struct btrfs_io_bio *btrfs_bio;
2717         struct bio *new;
2718
2719         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2720         if (new) {
2721                 btrfs_bio = btrfs_io_bio(new);
2722                 btrfs_bio->csum = NULL;
2723                 btrfs_bio->csum_allocated = NULL;
2724                 btrfs_bio->end_io = NULL;
2725         }
2726         return new;
2727 }
2728
2729 /* this also allocates from the btrfs_bioset */
2730 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2731 {
2732         struct btrfs_io_bio *btrfs_bio;
2733         struct bio *bio;
2734
2735         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2736         if (bio) {
2737                 btrfs_bio = btrfs_io_bio(bio);
2738                 btrfs_bio->csum = NULL;
2739                 btrfs_bio->csum_allocated = NULL;
2740                 btrfs_bio->end_io = NULL;
2741         }
2742         return bio;
2743 }
2744
2745
2746 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2747                                        unsigned long bio_flags)
2748 {
2749         blk_status_t ret = 0;
2750         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2751         struct page *page = bvec->bv_page;
2752         struct extent_io_tree *tree = bio->bi_private;
2753         u64 start;
2754
2755         start = page_offset(page) + bvec->bv_offset;
2756
2757         bio->bi_private = NULL;
2758         bio_get(bio);
2759
2760         if (tree->ops)
2761                 ret = tree->ops->submit_bio_hook(page->mapping->host, bio,
2762                                            mirror_num, bio_flags, start);
2763         else
2764                 btrfsic_submit_bio(bio);
2765
2766         bio_put(bio);
2767         return blk_status_to_errno(ret);
2768 }
2769
2770 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2771                      unsigned long offset, size_t size, struct bio *bio,
2772                      unsigned long bio_flags)
2773 {
2774         int ret = 0;
2775         if (tree->ops)
2776                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2777                                                 bio_flags);
2778         return ret;
2779
2780 }
2781
2782 static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree,
2783                               struct writeback_control *wbc,
2784                               struct page *page, sector_t sector,
2785                               size_t size, unsigned long offset,
2786                               struct block_device *bdev,
2787                               struct bio **bio_ret,
2788                               bio_end_io_t end_io_func,
2789                               int mirror_num,
2790                               unsigned long prev_bio_flags,
2791                               unsigned long bio_flags,
2792                               bool force_bio_submit)
2793 {
2794         int ret = 0;
2795         struct bio *bio;
2796         int contig = 0;
2797         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2798         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2799
2800         if (bio_ret && *bio_ret) {
2801                 bio = *bio_ret;
2802                 if (old_compressed)
2803                         contig = bio->bi_iter.bi_sector == sector;
2804                 else
2805                         contig = bio_end_sector(bio) == sector;
2806
2807                 if (prev_bio_flags != bio_flags || !contig ||
2808                     force_bio_submit ||
2809                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2810                     bio_add_page(bio, page, page_size, offset) < page_size) {
2811                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2812                         if (ret < 0) {
2813                                 *bio_ret = NULL;
2814                                 return ret;
2815                         }
2816                         bio = NULL;
2817                 } else {
2818                         if (wbc)
2819                                 wbc_account_io(wbc, page, page_size);
2820                         return 0;
2821                 }
2822         }
2823
2824         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2825                         GFP_NOFS | __GFP_HIGH);
2826         if (!bio)
2827                 return -ENOMEM;
2828
2829         bio_add_page(bio, page, page_size, offset);
2830         bio->bi_end_io = end_io_func;
2831         bio->bi_private = tree;
2832         bio_set_op_attrs(bio, op, op_flags);
2833         if (wbc) {
2834                 wbc_init_bio(wbc, bio);
2835                 wbc_account_io(wbc, page, page_size);
2836         }
2837
2838         if (bio_ret)
2839                 *bio_ret = bio;
2840         else
2841                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2842
2843         return ret;
2844 }
2845
2846 static void attach_extent_buffer_page(struct extent_buffer *eb,
2847                                       struct page *page)
2848 {
2849         if (!PagePrivate(page)) {
2850                 SetPagePrivate(page);
2851                 get_page(page);
2852                 set_page_private(page, (unsigned long)eb);
2853         } else {
2854                 WARN_ON(page->private != (unsigned long)eb);
2855         }
2856 }
2857
2858 void set_page_extent_mapped(struct page *page)
2859 {
2860         if (!PagePrivate(page)) {
2861                 SetPagePrivate(page);
2862                 get_page(page);
2863                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2864         }
2865 }
2866
2867 static struct extent_map *
2868 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2869                  u64 start, u64 len, get_extent_t *get_extent,
2870                  struct extent_map **em_cached)
2871 {
2872         struct extent_map *em;
2873
2874         if (em_cached && *em_cached) {
2875                 em = *em_cached;
2876                 if (extent_map_in_tree(em) && start >= em->start &&
2877                     start < extent_map_end(em)) {
2878                         refcount_inc(&em->refs);
2879                         return em;
2880                 }
2881
2882                 free_extent_map(em);
2883                 *em_cached = NULL;
2884         }
2885
2886         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2887         if (em_cached && !IS_ERR_OR_NULL(em)) {
2888                 BUG_ON(*em_cached);
2889                 refcount_inc(&em->refs);
2890                 *em_cached = em;
2891         }
2892         return em;
2893 }
2894 /*
2895  * basic readpage implementation.  Locked extent state structs are inserted
2896  * into the tree that are removed when the IO is done (by the end_io
2897  * handlers)
2898  * XXX JDM: This needs looking at to ensure proper page locking
2899  * return 0 on success, otherwise return error
2900  */
2901 static int __do_readpage(struct extent_io_tree *tree,
2902                          struct page *page,
2903                          get_extent_t *get_extent,
2904                          struct extent_map **em_cached,
2905                          struct bio **bio, int mirror_num,
2906                          unsigned long *bio_flags, int read_flags,
2907                          u64 *prev_em_start)
2908 {
2909         struct inode *inode = page->mapping->host;
2910         u64 start = page_offset(page);
2911         u64 page_end = start + PAGE_SIZE - 1;
2912         u64 end;
2913         u64 cur = start;
2914         u64 extent_offset;
2915         u64 last_byte = i_size_read(inode);
2916         u64 block_start;
2917         u64 cur_end;
2918         sector_t sector;
2919         struct extent_map *em;
2920         struct block_device *bdev;
2921         int ret = 0;
2922         int nr = 0;
2923         size_t pg_offset = 0;
2924         size_t iosize;
2925         size_t disk_io_size;
2926         size_t blocksize = inode->i_sb->s_blocksize;
2927         unsigned long this_bio_flag = 0;
2928
2929         set_page_extent_mapped(page);
2930
2931         end = page_end;
2932         if (!PageUptodate(page)) {
2933                 if (cleancache_get_page(page) == 0) {
2934                         BUG_ON(blocksize != PAGE_SIZE);
2935                         unlock_extent(tree, start, end);
2936                         goto out;
2937                 }
2938         }
2939
2940         if (page->index == last_byte >> PAGE_SHIFT) {
2941                 char *userpage;
2942                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2943
2944                 if (zero_offset) {
2945                         iosize = PAGE_SIZE - zero_offset;
2946                         userpage = kmap_atomic(page);
2947                         memset(userpage + zero_offset, 0, iosize);
2948                         flush_dcache_page(page);
2949                         kunmap_atomic(userpage);
2950                 }
2951         }
2952         while (cur <= end) {
2953                 bool force_bio_submit = false;
2954
2955                 if (cur >= last_byte) {
2956                         char *userpage;
2957                         struct extent_state *cached = NULL;
2958
2959                         iosize = PAGE_SIZE - pg_offset;
2960                         userpage = kmap_atomic(page);
2961                         memset(userpage + pg_offset, 0, iosize);
2962                         flush_dcache_page(page);
2963                         kunmap_atomic(userpage);
2964                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2965                                             &cached, GFP_NOFS);
2966                         unlock_extent_cached(tree, cur,
2967                                              cur + iosize - 1,
2968                                              &cached, GFP_NOFS);
2969                         break;
2970                 }
2971                 em = __get_extent_map(inode, page, pg_offset, cur,
2972                                       end - cur + 1, get_extent, em_cached);
2973                 if (IS_ERR_OR_NULL(em)) {
2974                         SetPageError(page);
2975                         unlock_extent(tree, cur, end);
2976                         break;
2977                 }
2978                 extent_offset = cur - em->start;
2979                 BUG_ON(extent_map_end(em) <= cur);
2980                 BUG_ON(end < cur);
2981
2982                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2983                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2984                         extent_set_compress_type(&this_bio_flag,
2985                                                  em->compress_type);
2986                 }
2987
2988                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2989                 cur_end = min(extent_map_end(em) - 1, end);
2990                 iosize = ALIGN(iosize, blocksize);
2991                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2992                         disk_io_size = em->block_len;
2993                         sector = em->block_start >> 9;
2994                 } else {
2995                         sector = (em->block_start + extent_offset) >> 9;
2996                         disk_io_size = iosize;
2997                 }
2998                 bdev = em->bdev;
2999                 block_start = em->block_start;
3000                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3001                         block_start = EXTENT_MAP_HOLE;
3002
3003                 /*
3004                  * If we have a file range that points to a compressed extent
3005                  * and it's followed by a consecutive file range that points to
3006                  * to the same compressed extent (possibly with a different
3007                  * offset and/or length, so it either points to the whole extent
3008                  * or only part of it), we must make sure we do not submit a
3009                  * single bio to populate the pages for the 2 ranges because
3010                  * this makes the compressed extent read zero out the pages
3011                  * belonging to the 2nd range. Imagine the following scenario:
3012                  *
3013                  *  File layout
3014                  *  [0 - 8K]                     [8K - 24K]
3015                  *    |                               |
3016                  *    |                               |
3017                  * points to extent X,         points to extent X,
3018                  * offset 4K, length of 8K     offset 0, length 16K
3019                  *
3020                  * [extent X, compressed length = 4K uncompressed length = 16K]
3021                  *
3022                  * If the bio to read the compressed extent covers both ranges,
3023                  * it will decompress extent X into the pages belonging to the
3024                  * first range and then it will stop, zeroing out the remaining
3025                  * pages that belong to the other range that points to extent X.
3026                  * So here we make sure we submit 2 bios, one for the first
3027                  * range and another one for the third range. Both will target
3028                  * the same physical extent from disk, but we can't currently
3029                  * make the compressed bio endio callback populate the pages
3030                  * for both ranges because each compressed bio is tightly
3031                  * coupled with a single extent map, and each range can have
3032                  * an extent map with a different offset value relative to the
3033                  * uncompressed data of our extent and different lengths. This
3034                  * is a corner case so we prioritize correctness over
3035                  * non-optimal behavior (submitting 2 bios for the same extent).
3036                  */
3037                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3038                     prev_em_start && *prev_em_start != (u64)-1 &&
3039                     *prev_em_start != em->orig_start)
3040                         force_bio_submit = true;
3041
3042                 if (prev_em_start)
3043                         *prev_em_start = em->orig_start;
3044
3045                 free_extent_map(em);
3046                 em = NULL;
3047
3048                 /* we've found a hole, just zero and go on */
3049                 if (block_start == EXTENT_MAP_HOLE) {
3050                         char *userpage;
3051                         struct extent_state *cached = NULL;
3052
3053                         userpage = kmap_atomic(page);
3054                         memset(userpage + pg_offset, 0, iosize);
3055                         flush_dcache_page(page);
3056                         kunmap_atomic(userpage);
3057
3058                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3059                                             &cached, GFP_NOFS);
3060                         unlock_extent_cached(tree, cur,
3061                                              cur + iosize - 1,
3062                                              &cached, GFP_NOFS);
3063                         cur = cur + iosize;
3064                         pg_offset += iosize;
3065                         continue;
3066                 }
3067                 /* the get_extent function already copied into the page */
3068                 if (test_range_bit(tree, cur, cur_end,
3069                                    EXTENT_UPTODATE, 1, NULL)) {
3070                         check_page_uptodate(tree, page);
3071                         unlock_extent(tree, cur, cur + iosize - 1);
3072                         cur = cur + iosize;
3073                         pg_offset += iosize;
3074                         continue;
3075                 }
3076                 /* we have an inline extent but it didn't get marked up
3077                  * to date.  Error out
3078                  */
3079                 if (block_start == EXTENT_MAP_INLINE) {
3080                         SetPageError(page);
3081                         unlock_extent(tree, cur, cur + iosize - 1);
3082                         cur = cur + iosize;
3083                         pg_offset += iosize;
3084                         continue;
3085                 }
3086
3087                 ret = submit_extent_page(REQ_OP_READ, read_flags, tree, NULL,
3088                                          page, sector, disk_io_size, pg_offset,
3089                                          bdev, bio,
3090                                          end_bio_extent_readpage, mirror_num,
3091                                          *bio_flags,
3092                                          this_bio_flag,
3093                                          force_bio_submit);
3094                 if (!ret) {
3095                         nr++;
3096                         *bio_flags = this_bio_flag;
3097                 } else {
3098                         SetPageError(page);
3099                         unlock_extent(tree, cur, cur + iosize - 1);
3100                         goto out;
3101                 }
3102                 cur = cur + iosize;
3103                 pg_offset += iosize;
3104         }
3105 out:
3106         if (!nr) {
3107                 if (!PageError(page))
3108                         SetPageUptodate(page);
3109                 unlock_page(page);
3110         }
3111         return ret;
3112 }
3113
3114 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3115                                              struct page *pages[], int nr_pages,
3116                                              u64 start, u64 end,
3117                                              get_extent_t *get_extent,
3118                                              struct extent_map **em_cached,
3119                                              struct bio **bio, int mirror_num,
3120                                              unsigned long *bio_flags,
3121                                              u64 *prev_em_start)
3122 {
3123         struct inode *inode;
3124         struct btrfs_ordered_extent *ordered;
3125         int index;
3126
3127         inode = pages[0]->mapping->host;
3128         while (1) {
3129                 lock_extent(tree, start, end);
3130                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3131                                                      end - start + 1);
3132                 if (!ordered)
3133                         break;
3134                 unlock_extent(tree, start, end);
3135                 btrfs_start_ordered_extent(inode, ordered, 1);
3136                 btrfs_put_ordered_extent(ordered);
3137         }
3138
3139         for (index = 0; index < nr_pages; index++) {
3140                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3141                               mirror_num, bio_flags, 0, prev_em_start);
3142                 put_page(pages[index]);
3143         }
3144 }
3145
3146 static void __extent_readpages(struct extent_io_tree *tree,
3147                                struct page *pages[],
3148                                int nr_pages, get_extent_t *get_extent,
3149                                struct extent_map **em_cached,
3150                                struct bio **bio, int mirror_num,
3151                                unsigned long *bio_flags,
3152                                u64 *prev_em_start)
3153 {
3154         u64 start = 0;
3155         u64 end = 0;
3156         u64 page_start;
3157         int index;
3158         int first_index = 0;
3159
3160         for (index = 0; index < nr_pages; index++) {
3161                 page_start = page_offset(pages[index]);
3162                 if (!end) {
3163                         start = page_start;
3164                         end = start + PAGE_SIZE - 1;
3165                         first_index = index;
3166                 } else if (end + 1 == page_start) {
3167                         end += PAGE_SIZE;
3168                 } else {
3169                         __do_contiguous_readpages(tree, &pages[first_index],
3170                                                   index - first_index, start,
3171                                                   end, get_extent, em_cached,
3172                                                   bio, mirror_num, bio_flags,
3173                                                   prev_em_start);
3174                         start = page_start;
3175                         end = start + PAGE_SIZE - 1;
3176                         first_index = index;
3177                 }
3178         }
3179
3180         if (end)
3181                 __do_contiguous_readpages(tree, &pages[first_index],
3182                                           index - first_index, start,
3183                                           end, get_extent, em_cached, bio,
3184                                           mirror_num, bio_flags,
3185                                           prev_em_start);
3186 }
3187
3188 static int __extent_read_full_page(struct extent_io_tree *tree,
3189                                    struct page *page,
3190                                    get_extent_t *get_extent,
3191                                    struct bio **bio, int mirror_num,
3192                                    unsigned long *bio_flags, int read_flags)
3193 {
3194         struct inode *inode = page->mapping->host;
3195         struct btrfs_ordered_extent *ordered;
3196         u64 start = page_offset(page);
3197         u64 end = start + PAGE_SIZE - 1;
3198         int ret;
3199
3200         while (1) {
3201                 lock_extent(tree, start, end);
3202                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3203                                                 PAGE_SIZE);
3204                 if (!ordered)
3205                         break;
3206                 unlock_extent(tree, start, end);
3207                 btrfs_start_ordered_extent(inode, ordered, 1);
3208                 btrfs_put_ordered_extent(ordered);
3209         }
3210
3211         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3212                             bio_flags, read_flags, NULL);
3213         return ret;
3214 }
3215
3216 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3217                             get_extent_t *get_extent, int mirror_num)
3218 {
3219         struct bio *bio = NULL;
3220         unsigned long bio_flags = 0;
3221         int ret;
3222
3223         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3224                                       &bio_flags, 0);
3225         if (bio)
3226                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3227         return ret;
3228 }
3229
3230 static void update_nr_written(struct writeback_control *wbc,
3231                               unsigned long nr_written)
3232 {
3233         wbc->nr_to_write -= nr_written;
3234 }
3235
3236 /*
3237  * helper for __extent_writepage, doing all of the delayed allocation setup.
3238  *
3239  * This returns 1 if our fill_delalloc function did all the work required
3240  * to write the page (copy into inline extent).  In this case the IO has
3241  * been started and the page is already unlocked.
3242  *
3243  * This returns 0 if all went well (page still locked)
3244  * This returns < 0 if there were errors (page still locked)
3245  */
3246 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3247                               struct page *page, struct writeback_control *wbc,
3248                               struct extent_page_data *epd,
3249                               u64 delalloc_start,
3250                               unsigned long *nr_written)
3251 {
3252         struct extent_io_tree *tree = epd->tree;
3253         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3254         u64 nr_delalloc;
3255         u64 delalloc_to_write = 0;
3256         u64 delalloc_end = 0;
3257         int ret;
3258         int page_started = 0;
3259
3260         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3261                 return 0;
3262
3263         while (delalloc_end < page_end) {
3264                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3265                                                page,
3266                                                &delalloc_start,
3267                                                &delalloc_end,
3268                                                BTRFS_MAX_EXTENT_SIZE);
3269                 if (nr_delalloc == 0) {
3270                         delalloc_start = delalloc_end + 1;
3271                         continue;
3272                 }
3273                 ret = tree->ops->fill_delalloc(inode, page,
3274                                                delalloc_start,
3275                                                delalloc_end,
3276                                                &page_started,
3277                                                nr_written);
3278                 /* File system has been set read-only */
3279                 if (ret) {
3280                         SetPageError(page);
3281                         /* fill_delalloc should be return < 0 for error
3282                          * but just in case, we use > 0 here meaning the
3283                          * IO is started, so we don't want to return > 0
3284                          * unless things are going well.
3285                          */
3286                         ret = ret < 0 ? ret : -EIO;
3287                         goto done;
3288                 }
3289                 /*
3290                  * delalloc_end is already one less than the total length, so
3291                  * we don't subtract one from PAGE_SIZE
3292                  */
3293                 delalloc_to_write += (delalloc_end - delalloc_start +
3294                                       PAGE_SIZE) >> PAGE_SHIFT;
3295                 delalloc_start = delalloc_end + 1;
3296         }
3297         if (wbc->nr_to_write < delalloc_to_write) {
3298                 int thresh = 8192;
3299
3300                 if (delalloc_to_write < thresh * 2)
3301                         thresh = delalloc_to_write;
3302                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3303                                          thresh);
3304         }
3305
3306         /* did the fill delalloc function already unlock and start
3307          * the IO?
3308          */
3309         if (page_started) {
3310                 /*
3311                  * we've unlocked the page, so we can't update
3312                  * the mapping's writeback index, just update
3313                  * nr_to_write.
3314                  */
3315                 wbc->nr_to_write -= *nr_written;
3316                 return 1;
3317         }
3318
3319         ret = 0;
3320
3321 done:
3322         return ret;
3323 }
3324
3325 /*
3326  * helper for __extent_writepage.  This calls the writepage start hooks,
3327  * and does the loop to map the page into extents and bios.
3328  *
3329  * We return 1 if the IO is started and the page is unlocked,
3330  * 0 if all went well (page still locked)
3331  * < 0 if there were errors (page still locked)
3332  */
3333 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3334                                  struct page *page,
3335                                  struct writeback_control *wbc,
3336                                  struct extent_page_data *epd,
3337                                  loff_t i_size,
3338                                  unsigned long nr_written,
3339                                  int write_flags, int *nr_ret)
3340 {
3341         struct extent_io_tree *tree = epd->tree;
3342         u64 start = page_offset(page);
3343         u64 page_end = start + PAGE_SIZE - 1;
3344         u64 end;
3345         u64 cur = start;
3346         u64 extent_offset;
3347         u64 block_start;
3348         u64 iosize;
3349         sector_t sector;
3350         struct extent_map *em;
3351         struct block_device *bdev;
3352         size_t pg_offset = 0;
3353         size_t blocksize;
3354         int ret = 0;
3355         int nr = 0;
3356         bool compressed;
3357
3358         if (tree->ops && tree->ops->writepage_start_hook) {
3359                 ret = tree->ops->writepage_start_hook(page, start,
3360                                                       page_end);
3361                 if (ret) {
3362                         /* Fixup worker will requeue */
3363                         if (ret == -EBUSY)
3364                                 wbc->pages_skipped++;
3365                         else
3366                                 redirty_page_for_writepage(wbc, page);
3367
3368                         update_nr_written(wbc, nr_written);
3369                         unlock_page(page);
3370                         return 1;
3371                 }
3372         }
3373
3374         /*
3375          * we don't want to touch the inode after unlocking the page,
3376          * so we update the mapping writeback index now
3377          */
3378         update_nr_written(wbc, nr_written + 1);
3379
3380         end = page_end;
3381         if (i_size <= start) {
3382                 if (tree->ops && tree->ops->writepage_end_io_hook)
3383                         tree->ops->writepage_end_io_hook(page, start,
3384                                                          page_end, NULL, 1);
3385                 goto done;
3386         }
3387
3388         blocksize = inode->i_sb->s_blocksize;
3389
3390         while (cur <= end) {
3391                 u64 em_end;
3392
3393                 if (cur >= i_size) {
3394                         if (tree->ops && tree->ops->writepage_end_io_hook)
3395                                 tree->ops->writepage_end_io_hook(page, cur,
3396                                                          page_end, NULL, 1);
3397                         break;
3398                 }
3399                 em = epd->get_extent(BTRFS_I(inode), page, pg_offset, cur,
3400                                      end - cur + 1, 1);
3401                 if (IS_ERR_OR_NULL(em)) {
3402                         SetPageError(page);
3403                         ret = PTR_ERR_OR_ZERO(em);
3404                         break;
3405                 }
3406
3407                 extent_offset = cur - em->start;
3408                 em_end = extent_map_end(em);
3409                 BUG_ON(em_end <= cur);
3410                 BUG_ON(end < cur);
3411                 iosize = min(em_end - cur, end - cur + 1);
3412                 iosize = ALIGN(iosize, blocksize);
3413                 sector = (em->block_start + extent_offset) >> 9;
3414                 bdev = em->bdev;
3415                 block_start = em->block_start;
3416                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3417                 free_extent_map(em);
3418                 em = NULL;
3419
3420                 /*
3421                  * compressed and inline extents are written through other
3422                  * paths in the FS
3423                  */
3424                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3425                     block_start == EXTENT_MAP_INLINE) {
3426                         /*
3427                          * end_io notification does not happen here for
3428                          * compressed extents
3429                          */
3430                         if (!compressed && tree->ops &&
3431                             tree->ops->writepage_end_io_hook)
3432                                 tree->ops->writepage_end_io_hook(page, cur,
3433                                                          cur + iosize - 1,
3434                                                          NULL, 1);
3435                         else if (compressed) {
3436                                 /* we don't want to end_page_writeback on
3437                                  * a compressed extent.  this happens
3438                                  * elsewhere
3439                                  */
3440                                 nr++;
3441                         }
3442
3443                         cur += iosize;
3444                         pg_offset += iosize;
3445                         continue;
3446                 }
3447
3448                 set_range_writeback(tree, cur, cur + iosize - 1);
3449                 if (!PageWriteback(page)) {
3450                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3451                                    "page %lu not writeback, cur %llu end %llu",
3452                                page->index, cur, end);
3453                 }
3454
3455                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3456                                          page, sector, iosize, pg_offset,
3457                                          bdev, &epd->bio,
3458                                          end_bio_extent_writepage,
3459                                          0, 0, 0, false);
3460                 if (ret) {
3461                         SetPageError(page);
3462                         if (PageWriteback(page))
3463                                 end_page_writeback(page);
3464                 }
3465
3466                 cur = cur + iosize;
3467                 pg_offset += iosize;
3468                 nr++;
3469         }
3470 done:
3471         *nr_ret = nr;
3472         return ret;
3473 }
3474
3475 /*
3476  * the writepage semantics are similar to regular writepage.  extent
3477  * records are inserted to lock ranges in the tree, and as dirty areas
3478  * are found, they are marked writeback.  Then the lock bits are removed
3479  * and the end_io handler clears the writeback ranges
3480  */
3481 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3482                               void *data)
3483 {
3484         struct inode *inode = page->mapping->host;
3485         struct extent_page_data *epd = data;
3486         u64 start = page_offset(page);
3487         u64 page_end = start + PAGE_SIZE - 1;
3488         int ret;
3489         int nr = 0;
3490         size_t pg_offset = 0;
3491         loff_t i_size = i_size_read(inode);
3492         unsigned long end_index = i_size >> PAGE_SHIFT;
3493         int write_flags = 0;
3494         unsigned long nr_written = 0;
3495
3496         if (wbc->sync_mode == WB_SYNC_ALL)
3497                 write_flags = REQ_SYNC;
3498
3499         trace___extent_writepage(page, inode, wbc);
3500
3501         WARN_ON(!PageLocked(page));
3502
3503         ClearPageError(page);
3504
3505         pg_offset = i_size & (PAGE_SIZE - 1);
3506         if (page->index > end_index ||
3507            (page->index == end_index && !pg_offset)) {
3508                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3509                 unlock_page(page);
3510                 return 0;
3511         }
3512
3513         if (page->index == end_index) {
3514                 char *userpage;
3515
3516                 userpage = kmap_atomic(page);
3517                 memset(userpage + pg_offset, 0,
3518                        PAGE_SIZE - pg_offset);
3519                 kunmap_atomic(userpage);
3520                 flush_dcache_page(page);
3521         }
3522
3523         pg_offset = 0;
3524
3525         set_page_extent_mapped(page);
3526
3527         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3528         if (ret == 1)
3529                 goto done_unlocked;
3530         if (ret)
3531                 goto done;
3532
3533         ret = __extent_writepage_io(inode, page, wbc, epd,
3534                                     i_size, nr_written, write_flags, &nr);
3535         if (ret == 1)
3536                 goto done_unlocked;
3537
3538 done:
3539         if (nr == 0) {
3540                 /* make sure the mapping tag for page dirty gets cleared */
3541                 set_page_writeback(page);
3542                 end_page_writeback(page);
3543         }
3544         if (PageError(page)) {
3545                 ret = ret < 0 ? ret : -EIO;
3546                 end_extent_writepage(page, ret, start, page_end);
3547         }
3548         unlock_page(page);
3549         return ret;
3550
3551 done_unlocked:
3552         return 0;
3553 }
3554
3555 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3556 {
3557         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3558                        TASK_UNINTERRUPTIBLE);
3559 }
3560
3561 static noinline_for_stack int
3562 lock_extent_buffer_for_io(struct extent_buffer *eb,
3563                           struct btrfs_fs_info *fs_info,
3564                           struct extent_page_data *epd)
3565 {
3566         unsigned long i, num_pages;
3567         int flush = 0;
3568         int ret = 0;
3569
3570         if (!btrfs_try_tree_write_lock(eb)) {
3571                 flush = 1;
3572                 flush_write_bio(epd);
3573                 btrfs_tree_lock(eb);
3574         }
3575
3576         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3577                 btrfs_tree_unlock(eb);
3578                 if (!epd->sync_io)
3579                         return 0;
3580                 if (!flush) {
3581                         flush_write_bio(epd);
3582                         flush = 1;
3583                 }
3584                 while (1) {
3585                         wait_on_extent_buffer_writeback(eb);
3586                         btrfs_tree_lock(eb);
3587                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3588                                 break;
3589                         btrfs_tree_unlock(eb);
3590                 }
3591         }
3592
3593         /*
3594          * We need to do this to prevent races in people who check if the eb is
3595          * under IO since we can end up having no IO bits set for a short period
3596          * of time.
3597          */
3598         spin_lock(&eb->refs_lock);
3599         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3600                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3601                 spin_unlock(&eb->refs_lock);
3602                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3603                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3604                                      -eb->len,
3605                                      fs_info->dirty_metadata_batch);
3606                 ret = 1;
3607         } else {
3608                 spin_unlock(&eb->refs_lock);
3609         }
3610
3611         btrfs_tree_unlock(eb);
3612
3613         if (!ret)
3614                 return ret;
3615
3616         num_pages = num_extent_pages(eb->start, eb->len);
3617         for (i = 0; i < num_pages; i++) {
3618                 struct page *p = eb->pages[i];
3619
3620                 if (!trylock_page(p)) {
3621                         if (!flush) {
3622                                 flush_write_bio(epd);
3623                                 flush = 1;
3624                         }
3625                         lock_page(p);
3626                 }
3627         }
3628
3629         return ret;
3630 }
3631
3632 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3633 {
3634         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3635         smp_mb__after_atomic();
3636         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3637 }
3638
3639 static void set_btree_ioerr(struct page *page)
3640 {
3641         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3642
3643         SetPageError(page);
3644         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3645                 return;
3646
3647         /*
3648          * If writeback for a btree extent that doesn't belong to a log tree
3649          * failed, increment the counter transaction->eb_write_errors.
3650          * We do this because while the transaction is running and before it's
3651          * committing (when we call filemap_fdata[write|wait]_range against
3652          * the btree inode), we might have
3653          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3654          * returns an error or an error happens during writeback, when we're
3655          * committing the transaction we wouldn't know about it, since the pages
3656          * can be no longer dirty nor marked anymore for writeback (if a
3657          * subsequent modification to the extent buffer didn't happen before the
3658          * transaction commit), which makes filemap_fdata[write|wait]_range not
3659          * able to find the pages tagged with SetPageError at transaction
3660          * commit time. So if this happens we must abort the transaction,
3661          * otherwise we commit a super block with btree roots that point to
3662          * btree nodes/leafs whose content on disk is invalid - either garbage
3663          * or the content of some node/leaf from a past generation that got
3664          * cowed or deleted and is no longer valid.
3665          *
3666          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3667          * not be enough - we need to distinguish between log tree extents vs
3668          * non-log tree extents, and the next filemap_fdatawait_range() call
3669          * will catch and clear such errors in the mapping - and that call might
3670          * be from a log sync and not from a transaction commit. Also, checking
3671          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3672          * not done and would not be reliable - the eb might have been released
3673          * from memory and reading it back again means that flag would not be
3674          * set (since it's a runtime flag, not persisted on disk).
3675          *
3676          * Using the flags below in the btree inode also makes us achieve the
3677          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3678          * writeback for all dirty pages and before filemap_fdatawait_range()
3679          * is called, the writeback for all dirty pages had already finished
3680          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3681          * filemap_fdatawait_range() would return success, as it could not know
3682          * that writeback errors happened (the pages were no longer tagged for
3683          * writeback).
3684          */
3685         switch (eb->log_index) {
3686         case -1:
3687                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3688                 break;
3689         case 0:
3690                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3691                 break;
3692         case 1:
3693                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3694                 break;
3695         default:
3696                 BUG(); /* unexpected, logic error */
3697         }
3698 }
3699
3700 static void end_bio_extent_buffer_writepage(struct bio *bio)
3701 {
3702         struct bio_vec *bvec;
3703         struct extent_buffer *eb;
3704         int i, done;
3705
3706         bio_for_each_segment_all(bvec, bio, i) {
3707                 struct page *page = bvec->bv_page;
3708
3709                 eb = (struct extent_buffer *)page->private;
3710                 BUG_ON(!eb);
3711                 done = atomic_dec_and_test(&eb->io_pages);
3712
3713                 if (bio->bi_status ||
3714                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3715                         ClearPageUptodate(page);
3716                         set_btree_ioerr(page);
3717                 }
3718
3719                 end_page_writeback(page);
3720
3721                 if (!done)
3722                         continue;
3723
3724                 end_extent_buffer_writeback(eb);
3725         }
3726
3727         bio_put(bio);
3728 }
3729
3730 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3731                         struct btrfs_fs_info *fs_info,
3732                         struct writeback_control *wbc,
3733                         struct extent_page_data *epd)
3734 {
3735         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3736         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3737         u64 offset = eb->start;
3738         u32 nritems;
3739         unsigned long i, num_pages;
3740         unsigned long bio_flags = 0;
3741         unsigned long start, end;
3742         int write_flags = (epd->sync_io ? REQ_SYNC : 0) | REQ_META;
3743         int ret = 0;
3744
3745         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3746         num_pages = num_extent_pages(eb->start, eb->len);
3747         atomic_set(&eb->io_pages, num_pages);
3748         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3749                 bio_flags = EXTENT_BIO_TREE_LOG;
3750
3751         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3752         nritems = btrfs_header_nritems(eb);
3753         if (btrfs_header_level(eb) > 0) {
3754                 end = btrfs_node_key_ptr_offset(nritems);
3755
3756                 memzero_extent_buffer(eb, end, eb->len - end);
3757         } else {
3758                 /*
3759                  * leaf:
3760                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3761                  */
3762                 start = btrfs_item_nr_offset(nritems);
3763                 end = btrfs_leaf_data(eb) + leaf_data_end(fs_info, eb);
3764                 memzero_extent_buffer(eb, start, end - start);
3765         }
3766
3767         for (i = 0; i < num_pages; i++) {
3768                 struct page *p = eb->pages[i];
3769
3770                 clear_page_dirty_for_io(p);
3771                 set_page_writeback(p);
3772                 ret = submit_extent_page(REQ_OP_WRITE, write_flags, tree, wbc,
3773                                          p, offset >> 9, PAGE_SIZE, 0, bdev,
3774                                          &epd->bio,
3775                                          end_bio_extent_buffer_writepage,
3776                                          0, epd->bio_flags, bio_flags, false);
3777                 epd->bio_flags = bio_flags;
3778                 if (ret) {
3779                         set_btree_ioerr(p);
3780                         if (PageWriteback(p))
3781                                 end_page_writeback(p);
3782                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3783                                 end_extent_buffer_writeback(eb);
3784                         ret = -EIO;
3785                         break;
3786                 }
3787                 offset += PAGE_SIZE;
3788                 update_nr_written(wbc, 1);
3789                 unlock_page(p);
3790         }
3791
3792         if (unlikely(ret)) {
3793                 for (; i < num_pages; i++) {
3794                         struct page *p = eb->pages[i];
3795                         clear_page_dirty_for_io(p);
3796                         unlock_page(p);
3797                 }
3798         }
3799
3800         return ret;
3801 }
3802
3803 int btree_write_cache_pages(struct address_space *mapping,
3804                                    struct writeback_control *wbc)
3805 {
3806         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3807         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3808         struct extent_buffer *eb, *prev_eb = NULL;
3809         struct extent_page_data epd = {
3810                 .bio = NULL,
3811                 .tree = tree,
3812                 .extent_locked = 0,
3813                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3814                 .bio_flags = 0,
3815         };
3816         int ret = 0;
3817         int done = 0;
3818         int nr_to_write_done = 0;
3819         struct pagevec pvec;
3820         int nr_pages;
3821         pgoff_t index;
3822         pgoff_t end;            /* Inclusive */
3823         int scanned = 0;
3824         int tag;
3825
3826         pagevec_init(&pvec, 0);
3827         if (wbc->range_cyclic) {
3828                 index = mapping->writeback_index; /* Start from prev offset */
3829                 end = -1;
3830         } else {
3831                 index = wbc->range_start >> PAGE_SHIFT;
3832                 end = wbc->range_end >> PAGE_SHIFT;
3833                 scanned = 1;
3834         }
3835         if (wbc->sync_mode == WB_SYNC_ALL)
3836                 tag = PAGECACHE_TAG_TOWRITE;
3837         else
3838                 tag = PAGECACHE_TAG_DIRTY;
3839 retry:
3840         if (wbc->sync_mode == WB_SYNC_ALL)
3841                 tag_pages_for_writeback(mapping, index, end);
3842         while (!done && !nr_to_write_done && (index <= end) &&
3843                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3844                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3845                 unsigned i;
3846
3847                 scanned = 1;
3848                 for (i = 0; i < nr_pages; i++) {
3849                         struct page *page = pvec.pages[i];
3850
3851                         if (!PagePrivate(page))
3852                                 continue;
3853
3854                         if (!wbc->range_cyclic && page->index > end) {
3855                                 done = 1;
3856                                 break;
3857                         }
3858
3859                         spin_lock(&mapping->private_lock);
3860                         if (!PagePrivate(page)) {
3861                                 spin_unlock(&mapping->private_lock);
3862                                 continue;
3863                         }
3864
3865                         eb = (struct extent_buffer *)page->private;
3866
3867                         /*
3868                          * Shouldn't happen and normally this would be a BUG_ON
3869                          * but no sense in crashing the users box for something
3870                          * we can survive anyway.
3871                          */
3872                         if (WARN_ON(!eb)) {
3873                                 spin_unlock(&mapping->private_lock);
3874                                 continue;
3875                         }
3876
3877                         if (eb == prev_eb) {
3878                                 spin_unlock(&mapping->private_lock);
3879                                 continue;
3880                         }
3881
3882                         ret = atomic_inc_not_zero(&eb->refs);
3883                         spin_unlock(&mapping->private_lock);
3884                         if (!ret)
3885                                 continue;
3886
3887                         prev_eb = eb;
3888                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3889                         if (!ret) {
3890                                 free_extent_buffer(eb);
3891                                 continue;
3892                         }
3893
3894                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3895                         if (ret) {
3896                                 done = 1;
3897                                 free_extent_buffer(eb);
3898                                 break;
3899                         }
3900                         free_extent_buffer(eb);
3901
3902                         /*
3903                          * the filesystem may choose to bump up nr_to_write.
3904                          * We have to make sure to honor the new nr_to_write
3905                          * at any time
3906                          */
3907                         nr_to_write_done = wbc->nr_to_write <= 0;
3908                 }
3909                 pagevec_release(&pvec);
3910                 cond_resched();
3911         }
3912         if (!scanned && !done) {
3913                 /*
3914                  * We hit the last page and there is more work to be done: wrap
3915                  * back to the start of the file
3916                  */
3917                 scanned = 1;
3918                 index = 0;
3919                 goto retry;
3920         }
3921         flush_write_bio(&epd);
3922         return ret;
3923 }
3924
3925 /**
3926  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3927  * @mapping: address space structure to write
3928  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3929  * @writepage: function called for each page
3930  * @data: data passed to writepage function
3931  *
3932  * If a page is already under I/O, write_cache_pages() skips it, even
3933  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3934  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3935  * and msync() need to guarantee that all the data which was dirty at the time
3936  * the call was made get new I/O started against them.  If wbc->sync_mode is
3937  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3938  * existing IO to complete.
3939  */
3940 static int extent_write_cache_pages(struct address_space *mapping,
3941                              struct writeback_control *wbc,
3942                              writepage_t writepage, void *data,
3943                              void (*flush_fn)(void *))
3944 {
3945         struct inode *inode = mapping->host;
3946         int ret = 0;
3947         int done = 0;
3948         int nr_to_write_done = 0;
3949         struct pagevec pvec;
3950         int nr_pages;
3951         pgoff_t index;
3952         pgoff_t end;            /* Inclusive */
3953         pgoff_t done_index;
3954         int range_whole = 0;
3955         int scanned = 0;
3956         int tag;
3957
3958         /*
3959          * We have to hold onto the inode so that ordered extents can do their
3960          * work when the IO finishes.  The alternative to this is failing to add
3961          * an ordered extent if the igrab() fails there and that is a huge pain
3962          * to deal with, so instead just hold onto the inode throughout the
3963          * writepages operation.  If it fails here we are freeing up the inode
3964          * anyway and we'd rather not waste our time writing out stuff that is
3965          * going to be truncated anyway.
3966          */
3967         if (!igrab(inode))
3968                 return 0;
3969
3970         pagevec_init(&pvec, 0);
3971         if (wbc->range_cyclic) {
3972                 index = mapping->writeback_index; /* Start from prev offset */
3973                 end = -1;
3974         } else {
3975                 index = wbc->range_start >> PAGE_SHIFT;
3976                 end = wbc->range_end >> PAGE_SHIFT;
3977                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3978                         range_whole = 1;
3979                 scanned = 1;
3980         }
3981         if (wbc->sync_mode == WB_SYNC_ALL)
3982                 tag = PAGECACHE_TAG_TOWRITE;
3983         else
3984                 tag = PAGECACHE_TAG_DIRTY;
3985 retry:
3986         if (wbc->sync_mode == WB_SYNC_ALL)
3987                 tag_pages_for_writeback(mapping, index, end);
3988         done_index = index;
3989         while (!done && !nr_to_write_done && (index <= end) &&
3990                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3991                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3992                 unsigned i;
3993
3994                 scanned = 1;
3995                 for (i = 0; i < nr_pages; i++) {
3996                         struct page *page = pvec.pages[i];
3997
3998                         done_index = page->index;
3999                         /*
4000                          * At this point we hold neither mapping->tree_lock nor
4001                          * lock on the page itself: the page may be truncated or
4002                          * invalidated (changing page->mapping to NULL), or even
4003                          * swizzled back from swapper_space to tmpfs file
4004                          * mapping
4005                          */
4006                         if (!trylock_page(page)) {
4007                                 flush_fn(data);
4008                                 lock_page(page);
4009                         }
4010
4011                         if (unlikely(page->mapping != mapping)) {
4012                                 unlock_page(page);
4013                                 continue;
4014                         }
4015
4016                         if (!wbc->range_cyclic && page->index > end) {
4017                                 done = 1;
4018                                 unlock_page(page);
4019                                 continue;
4020                         }
4021
4022                         if (wbc->sync_mode != WB_SYNC_NONE) {
4023                                 if (PageWriteback(page))
4024                                         flush_fn(data);
4025                                 wait_on_page_writeback(page);
4026                         }
4027
4028                         if (PageWriteback(page) ||
4029                             !clear_page_dirty_for_io(page)) {
4030                                 unlock_page(page);
4031                                 continue;
4032                         }
4033
4034                         ret = (*writepage)(page, wbc, data);
4035
4036                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
4037                                 unlock_page(page);
4038                                 ret = 0;
4039                         }
4040                         if (ret < 0) {
4041                                 /*
4042                                  * done_index is set past this page,
4043                                  * so media errors will not choke
4044                                  * background writeout for the entire
4045                                  * file. This has consequences for
4046                                  * range_cyclic semantics (ie. it may
4047                                  * not be suitable for data integrity
4048                                  * writeout).
4049                                  */
4050                                 done_index = page->index + 1;
4051                                 done = 1;
4052                                 break;
4053                         }
4054
4055                         /*
4056                          * the filesystem may choose to bump up nr_to_write.
4057                          * We have to make sure to honor the new nr_to_write
4058                          * at any time
4059                          */
4060                         nr_to_write_done = wbc->nr_to_write <= 0;
4061                 }
4062                 pagevec_release(&pvec);
4063                 cond_resched();
4064         }
4065         if (!scanned && !done) {
4066                 /*
4067                  * We hit the last page and there is more work to be done: wrap
4068                  * back to the start of the file
4069                  */
4070                 scanned = 1;
4071                 index = 0;
4072                 goto retry;
4073         }
4074
4075         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4076                 mapping->writeback_index = done_index;
4077
4078         btrfs_add_delayed_iput(inode);
4079         return ret;
4080 }
4081
4082 static void flush_epd_write_bio(struct extent_page_data *epd)
4083 {
4084         if (epd->bio) {
4085                 int ret;
4086
4087                 bio_set_op_attrs(epd->bio, REQ_OP_WRITE,
4088                                  epd->sync_io ? REQ_SYNC : 0);
4089
4090                 ret = submit_one_bio(epd->bio, 0, epd->bio_flags);
4091                 BUG_ON(ret < 0); /* -ENOMEM */
4092                 epd->bio = NULL;
4093         }
4094 }
4095
4096 static noinline void flush_write_bio(void *data)
4097 {
4098         struct extent_page_data *epd = data;
4099         flush_epd_write_bio(epd);
4100 }
4101
4102 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
4103                           get_extent_t *get_extent,
4104                           struct writeback_control *wbc)
4105 {
4106         int ret;
4107         struct extent_page_data epd = {
4108                 .bio = NULL,
4109                 .tree = tree,
4110                 .get_extent = get_extent,
4111                 .extent_locked = 0,
4112                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4113                 .bio_flags = 0,
4114         };
4115
4116         ret = __extent_writepage(page, wbc, &epd);
4117
4118         flush_epd_write_bio(&epd);
4119         return ret;
4120 }
4121
4122 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
4123                               u64 start, u64 end, get_extent_t *get_extent,
4124                               int mode)
4125 {
4126         int ret = 0;
4127         struct address_space *mapping = inode->i_mapping;
4128         struct page *page;
4129         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4130                 PAGE_SHIFT;
4131
4132         struct extent_page_data epd = {
4133                 .bio = NULL,
4134                 .tree = tree,
4135                 .get_extent = get_extent,
4136                 .extent_locked = 1,
4137                 .sync_io = mode == WB_SYNC_ALL,
4138                 .bio_flags = 0,
4139         };
4140         struct writeback_control wbc_writepages = {
4141                 .sync_mode      = mode,
4142                 .nr_to_write    = nr_pages * 2,
4143                 .range_start    = start,
4144                 .range_end      = end + 1,
4145         };
4146
4147         while (start <= end) {
4148                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4149                 if (clear_page_dirty_for_io(page))
4150                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4151                 else {
4152                         if (tree->ops && tree->ops->writepage_end_io_hook)
4153                                 tree->ops->writepage_end_io_hook(page, start,
4154                                                  start + PAGE_SIZE - 1,
4155                                                  NULL, 1);
4156                         unlock_page(page);
4157                 }
4158                 put_page(page);
4159                 start += PAGE_SIZE;
4160         }
4161
4162         flush_epd_write_bio(&epd);
4163         return ret;
4164 }
4165
4166 int extent_writepages(struct extent_io_tree *tree,
4167                       struct address_space *mapping,
4168                       get_extent_t *get_extent,
4169                       struct writeback_control *wbc)
4170 {
4171         int ret = 0;
4172         struct extent_page_data epd = {
4173                 .bio = NULL,
4174                 .tree = tree,
4175                 .get_extent = get_extent,
4176                 .extent_locked = 0,
4177                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4178                 .bio_flags = 0,
4179         };
4180
4181         ret = extent_write_cache_pages(mapping, wbc, __extent_writepage, &epd,
4182                                        flush_write_bio);
4183         flush_epd_write_bio(&epd);
4184         return ret;
4185 }
4186
4187 int extent_readpages(struct extent_io_tree *tree,
4188                      struct address_space *mapping,
4189                      struct list_head *pages, unsigned nr_pages,
4190                      get_extent_t get_extent)
4191 {
4192         struct bio *bio = NULL;
4193         unsigned page_idx;
4194         unsigned long bio_flags = 0;
4195         struct page *pagepool[16];
4196         struct page *page;
4197         struct extent_map *em_cached = NULL;
4198         int nr = 0;
4199         u64 prev_em_start = (u64)-1;
4200
4201         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4202                 page = list_entry(pages->prev, struct page, lru);
4203
4204                 prefetchw(&page->flags);
4205                 list_del(&page->lru);
4206                 if (add_to_page_cache_lru(page, mapping,
4207                                         page->index,
4208                                         readahead_gfp_mask(mapping))) {
4209                         put_page(page);
4210                         continue;
4211                 }
4212
4213                 pagepool[nr++] = page;
4214                 if (nr < ARRAY_SIZE(pagepool))
4215                         continue;
4216                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4217                                    &bio, 0, &bio_flags, &prev_em_start);
4218                 nr = 0;
4219         }
4220         if (nr)
4221                 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
4222                                    &bio, 0, &bio_flags, &prev_em_start);
4223
4224         if (em_cached)
4225                 free_extent_map(em_cached);
4226
4227         BUG_ON(!list_empty(pages));
4228         if (bio)
4229                 return submit_one_bio(bio, 0, bio_flags);
4230         return 0;
4231 }
4232
4233 /*
4234  * basic invalidatepage code, this waits on any locked or writeback
4235  * ranges corresponding to the page, and then deletes any extent state
4236  * records from the tree
4237  */
4238 int extent_invalidatepage(struct extent_io_tree *tree,
4239                           struct page *page, unsigned long offset)
4240 {
4241         struct extent_state *cached_state = NULL;
4242         u64 start = page_offset(page);
4243         u64 end = start + PAGE_SIZE - 1;
4244         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4245
4246         start += ALIGN(offset, blocksize);
4247         if (start > end)
4248                 return 0;
4249
4250         lock_extent_bits(tree, start, end, &cached_state);
4251         wait_on_page_writeback(page);
4252         clear_extent_bit(tree, start, end,
4253                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4254                          EXTENT_DO_ACCOUNTING,
4255                          1, 1, &cached_state, GFP_NOFS);
4256         return 0;
4257 }
4258
4259 /*
4260  * a helper for releasepage, this tests for areas of the page that
4261  * are locked or under IO and drops the related state bits if it is safe
4262  * to drop the page.
4263  */
4264 static int try_release_extent_state(struct extent_map_tree *map,
4265                                     struct extent_io_tree *tree,
4266                                     struct page *page, gfp_t mask)
4267 {
4268         u64 start = page_offset(page);
4269         u64 end = start + PAGE_SIZE - 1;
4270         int ret = 1;
4271
4272         if (test_range_bit(tree, start, end,
4273                            EXTENT_IOBITS, 0, NULL))
4274                 ret = 0;
4275         else {
4276                 /*
4277                  * at this point we can safely clear everything except the
4278                  * locked bit and the nodatasum bit
4279                  */
4280                 ret = clear_extent_bit(tree, start, end,
4281                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4282                                  0, 0, NULL, mask);
4283
4284                 /* if clear_extent_bit failed for enomem reasons,
4285                  * we can't allow the release to continue.
4286                  */
4287                 if (ret < 0)
4288                         ret = 0;
4289                 else
4290                         ret = 1;
4291         }
4292         return ret;
4293 }
4294
4295 /*
4296  * a helper for releasepage.  As long as there are no locked extents
4297  * in the range corresponding to the page, both state records and extent
4298  * map records are removed
4299  */
4300 int try_release_extent_mapping(struct extent_map_tree *map,
4301                                struct extent_io_tree *tree, struct page *page,
4302                                gfp_t mask)
4303 {
4304         struct extent_map *em;
4305         u64 start = page_offset(page);
4306         u64 end = start + PAGE_SIZE - 1;
4307
4308         if (gfpflags_allow_blocking(mask) &&
4309             page->mapping->host->i_size > SZ_16M) {
4310                 u64 len;
4311                 while (start <= end) {
4312                         len = end - start + 1;
4313                         write_lock(&map->lock);
4314                         em = lookup_extent_mapping(map, start, len);
4315                         if (!em) {
4316                                 write_unlock(&map->lock);
4317                                 break;
4318                         }
4319                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4320                             em->start != start) {
4321                                 write_unlock(&map->lock);
4322                                 free_extent_map(em);
4323                                 break;
4324                         }
4325                         if (!test_range_bit(tree, em->start,
4326                                             extent_map_end(em) - 1,
4327                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4328                                             0, NULL)) {
4329                                 remove_extent_mapping(map, em);
4330                                 /* once for the rb tree */
4331                                 free_extent_map(em);
4332                         }
4333                         start = extent_map_end(em);
4334                         write_unlock(&map->lock);
4335
4336                         /* once for us */
4337                         free_extent_map(em);
4338                 }
4339         }
4340         return try_release_extent_state(map, tree, page, mask);
4341 }
4342
4343 /*
4344  * helper function for fiemap, which doesn't want to see any holes.
4345  * This maps until we find something past 'last'
4346  */
4347 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4348                                                 u64 offset,
4349                                                 u64 last,
4350                                                 get_extent_t *get_extent)
4351 {
4352         u64 sectorsize = btrfs_inode_sectorsize(inode);
4353         struct extent_map *em;
4354         u64 len;
4355
4356         if (offset >= last)
4357                 return NULL;
4358
4359         while (1) {
4360                 len = last - offset;
4361                 if (len == 0)
4362                         break;
4363                 len = ALIGN(len, sectorsize);
4364                 em = get_extent(BTRFS_I(inode), NULL, 0, offset, len, 0);
4365                 if (IS_ERR_OR_NULL(em))
4366                         return em;
4367
4368                 /* if this isn't a hole return it */
4369                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4370                     em->block_start != EXTENT_MAP_HOLE) {
4371                         return em;
4372                 }
4373
4374                 /* this is a hole, advance to the next extent */
4375                 offset = extent_map_end(em);
4376                 free_extent_map(em);
4377                 if (offset >= last)
4378                         break;
4379         }
4380         return NULL;
4381 }
4382
4383 /*
4384  * To cache previous fiemap extent
4385  *
4386  * Will be used for merging fiemap extent
4387  */
4388 struct fiemap_cache {
4389         u64 offset;
4390         u64 phys;
4391         u64 len;
4392         u32 flags;
4393         bool cached;
4394 };
4395
4396 /*
4397  * Helper to submit fiemap extent.
4398  *
4399  * Will try to merge current fiemap extent specified by @offset, @phys,
4400  * @len and @flags with cached one.
4401  * And only when we fails to merge, cached one will be submitted as
4402  * fiemap extent.
4403  *
4404  * Return value is the same as fiemap_fill_next_extent().
4405  */
4406 static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
4407                                 struct fiemap_cache *cache,
4408                                 u64 offset, u64 phys, u64 len, u32 flags)
4409 {
4410         int ret = 0;
4411
4412         if (!cache->cached)
4413                 goto assign;
4414
4415         /*
4416          * Sanity check, extent_fiemap() should have ensured that new
4417          * fiemap extent won't overlap with cahced one.
4418          * Not recoverable.
4419          *
4420          * NOTE: Physical address can overlap, due to compression
4421          */
4422         if (cache->offset + cache->len > offset) {
4423                 WARN_ON(1);
4424                 return -EINVAL;
4425         }
4426
4427         /*
4428          * Only merges fiemap extents if
4429          * 1) Their logical addresses are continuous
4430          *
4431          * 2) Their physical addresses are continuous
4432          *    So truly compressed (physical size smaller than logical size)
4433          *    extents won't get merged with each other
4434          *
4435          * 3) Share same flags except FIEMAP_EXTENT_LAST
4436          *    So regular extent won't get merged with prealloc extent
4437          */
4438         if (cache->offset + cache->len  == offset &&
4439             cache->phys + cache->len == phys  &&
4440             (cache->flags & ~FIEMAP_EXTENT_LAST) ==
4441                         (flags & ~FIEMAP_EXTENT_LAST)) {
4442                 cache->len += len;
4443                 cache->flags |= flags;
4444                 goto try_submit_last;
4445         }
4446
4447         /* Not mergeable, need to submit cached one */
4448         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4449                                       cache->len, cache->flags);
4450         cache->cached = false;
4451         if (ret)
4452                 return ret;
4453 assign:
4454         cache->cached = true;
4455         cache->offset = offset;
4456         cache->phys = phys;
4457         cache->len = len;
4458         cache->flags = flags;
4459 try_submit_last:
4460         if (cache->flags & FIEMAP_EXTENT_LAST) {
4461                 ret = fiemap_fill_next_extent(fieinfo, cache->offset,
4462                                 cache->phys, cache->len, cache->flags);
4463                 cache->cached = false;
4464         }
4465         return ret;
4466 }
4467
4468 /*
4469  * Sanity check for fiemap cache
4470  *
4471  * All fiemap cache should be submitted by emit_fiemap_extent()
4472  * Iteration should be terminated either by last fiemap extent or
4473  * fieinfo->fi_extents_max.
4474  * So no cached fiemap should exist.
4475  */
4476 static int check_fiemap_cache(struct btrfs_fs_info *fs_info,
4477                                struct fiemap_extent_info *fieinfo,
4478                                struct fiemap_cache *cache)
4479 {
4480         int ret;
4481
4482         if (!cache->cached)
4483                 return 0;
4484
4485         /* Small and recoverbale problem, only to info developer */
4486 #ifdef CONFIG_BTRFS_DEBUG
4487         WARN_ON(1);
4488 #endif
4489         btrfs_warn(fs_info,
4490                    "unhandled fiemap cache detected: offset=%llu phys=%llu len=%llu flags=0x%x",
4491                    cache->offset, cache->phys, cache->len, cache->flags);
4492         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4493                                       cache->len, cache->flags);
4494         cache->cached = false;
4495         if (ret > 0)
4496                 ret = 0;
4497         return ret;
4498 }
4499
4500 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4501                 __u64 start, __u64 len, get_extent_t *get_extent)
4502 {
4503         int ret = 0;
4504         u64 off = start;
4505         u64 max = start + len;
4506         u32 flags = 0;
4507         u32 found_type;
4508         u64 last;
4509         u64 last_for_get_extent = 0;
4510         u64 disko = 0;
4511         u64 isize = i_size_read(inode);
4512         struct btrfs_key found_key;
4513         struct extent_map *em = NULL;
4514         struct extent_state *cached_state = NULL;
4515         struct btrfs_path *path;
4516         struct btrfs_root *root = BTRFS_I(inode)->root;
4517         struct fiemap_cache cache = { 0 };
4518         int end = 0;
4519         u64 em_start = 0;
4520         u64 em_len = 0;
4521         u64 em_end = 0;
4522
4523         if (len == 0)
4524                 return -EINVAL;
4525
4526         path = btrfs_alloc_path();
4527         if (!path)
4528                 return -ENOMEM;
4529         path->leave_spinning = 1;
4530
4531         start = round_down(start, btrfs_inode_sectorsize(inode));
4532         len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
4533
4534         /*
4535          * lookup the last file extent.  We're not using i_size here
4536          * because there might be preallocation past i_size
4537          */
4538         ret = btrfs_lookup_file_extent(NULL, root, path,
4539                         btrfs_ino(BTRFS_I(inode)), -1, 0);
4540         if (ret < 0) {
4541                 btrfs_free_path(path);
4542                 return ret;
4543         } else {
4544                 WARN_ON(!ret);
4545                 if (ret == 1)
4546                         ret = 0;
4547         }
4548
4549         path->slots[0]--;
4550         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4551         found_type = found_key.type;
4552
4553         /* No extents, but there might be delalloc bits */
4554         if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) ||
4555             found_type != BTRFS_EXTENT_DATA_KEY) {
4556                 /* have to trust i_size as the end */
4557                 last = (u64)-1;
4558                 last_for_get_extent = isize;
4559         } else {
4560                 /*
4561                  * remember the start of the last extent.  There are a
4562                  * bunch of different factors that go into the length of the
4563                  * extent, so its much less complex to remember where it started
4564                  */
4565                 last = found_key.offset;
4566                 last_for_get_extent = last + 1;
4567         }
4568         btrfs_release_path(path);
4569
4570         /*
4571          * we might have some extents allocated but more delalloc past those
4572          * extents.  so, we trust isize unless the start of the last extent is
4573          * beyond isize
4574          */
4575         if (last < isize) {
4576                 last = (u64)-1;
4577                 last_for_get_extent = isize;
4578         }
4579
4580         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4581                          &cached_state);
4582
4583         em = get_extent_skip_holes(inode, start, last_for_get_extent,
4584                                    get_extent);
4585         if (!em)
4586                 goto out;
4587         if (IS_ERR(em)) {
4588                 ret = PTR_ERR(em);
4589                 goto out;
4590         }
4591
4592         while (!end) {
4593                 u64 offset_in_extent = 0;
4594
4595                 /* break if the extent we found is outside the range */
4596                 if (em->start >= max || extent_map_end(em) < off)
4597                         break;
4598
4599                 /*
4600                  * get_extent may return an extent that starts before our
4601                  * requested range.  We have to make sure the ranges
4602                  * we return to fiemap always move forward and don't
4603                  * overlap, so adjust the offsets here
4604                  */
4605                 em_start = max(em->start, off);
4606
4607                 /*
4608                  * record the offset from the start of the extent
4609                  * for adjusting the disk offset below.  Only do this if the
4610                  * extent isn't compressed since our in ram offset may be past
4611                  * what we have actually allocated on disk.
4612                  */
4613                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4614                         offset_in_extent = em_start - em->start;
4615                 em_end = extent_map_end(em);
4616                 em_len = em_end - em_start;
4617                 disko = 0;
4618                 flags = 0;
4619
4620                 /*
4621                  * bump off for our next call to get_extent
4622                  */
4623                 off = extent_map_end(em);
4624                 if (off >= max)
4625                         end = 1;
4626
4627                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4628                         end = 1;
4629                         flags |= FIEMAP_EXTENT_LAST;
4630                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4631                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4632                                   FIEMAP_EXTENT_NOT_ALIGNED);
4633                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4634                         flags |= (FIEMAP_EXTENT_DELALLOC |
4635                                   FIEMAP_EXTENT_UNKNOWN);
4636                 } else if (fieinfo->fi_extents_max) {
4637                         struct btrfs_trans_handle *trans;
4638
4639                         u64 bytenr = em->block_start -
4640                                 (em->start - em->orig_start);
4641
4642                         disko = em->block_start + offset_in_extent;
4643
4644                         /*
4645                          * We need a trans handle to get delayed refs
4646                          */
4647                         trans = btrfs_join_transaction(root);
4648                         /*
4649                          * It's OK if we can't start a trans we can still check
4650                          * from commit_root
4651                          */
4652                         if (IS_ERR(trans))
4653                                 trans = NULL;
4654
4655                         /*
4656                          * As btrfs supports shared space, this information
4657                          * can be exported to userspace tools via
4658                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4659                          * then we're just getting a count and we can skip the
4660                          * lookup stuff.
4661                          */
4662                         ret = btrfs_check_shared(trans, root->fs_info,
4663                                         root->objectid,
4664                                         btrfs_ino(BTRFS_I(inode)), bytenr);
4665                         if (trans)
4666                                 btrfs_end_transaction(trans);
4667                         if (ret < 0)
4668                                 goto out_free;
4669                         if (ret)
4670                                 flags |= FIEMAP_EXTENT_SHARED;
4671                         ret = 0;
4672                 }
4673                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4674                         flags |= FIEMAP_EXTENT_ENCODED;
4675                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4676                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4677
4678                 free_extent_map(em);
4679                 em = NULL;
4680                 if ((em_start >= last) || em_len == (u64)-1 ||
4681                    (last == (u64)-1 && isize <= em_end)) {
4682                         flags |= FIEMAP_EXTENT_LAST;
4683                         end = 1;
4684                 }
4685
4686                 /* now scan forward to see if this is really the last extent. */
4687                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4688                                            get_extent);
4689                 if (IS_ERR(em)) {
4690                         ret = PTR_ERR(em);
4691                         goto out;
4692                 }
4693                 if (!em) {
4694                         flags |= FIEMAP_EXTENT_LAST;
4695                         end = 1;
4696                 }
4697                 ret = emit_fiemap_extent(fieinfo, &cache, em_start, disko,
4698                                            em_len, flags);
4699                 if (ret) {
4700                         if (ret == 1)
4701                                 ret = 0;
4702                         goto out_free;
4703                 }
4704         }
4705 out_free:
4706         if (!ret)
4707                 ret = check_fiemap_cache(root->fs_info, fieinfo, &cache);
4708         free_extent_map(em);
4709 out:
4710         btrfs_free_path(path);
4711         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4712                              &cached_state, GFP_NOFS);
4713         return ret;
4714 }
4715
4716 static void __free_extent_buffer(struct extent_buffer *eb)
4717 {
4718         btrfs_leak_debug_del(&eb->leak_list);
4719         kmem_cache_free(extent_buffer_cache, eb);
4720 }
4721
4722 int extent_buffer_under_io(struct extent_buffer *eb)
4723 {
4724         return (atomic_read(&eb->io_pages) ||
4725                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4726                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4727 }
4728
4729 /*
4730  * Helper for releasing extent buffer page.
4731  */
4732 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4733 {
4734         unsigned long index;
4735         struct page *page;
4736         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4737
4738         BUG_ON(extent_buffer_under_io(eb));
4739
4740         index = num_extent_pages(eb->start, eb->len);
4741         if (index == 0)
4742                 return;
4743
4744         do {
4745                 index--;
4746                 page = eb->pages[index];
4747                 if (!page)
4748                         continue;
4749                 if (mapped)
4750                         spin_lock(&page->mapping->private_lock);
4751                 /*
4752                  * We do this since we'll remove the pages after we've
4753                  * removed the eb from the radix tree, so we could race
4754                  * and have this page now attached to the new eb.  So
4755                  * only clear page_private if it's still connected to
4756                  * this eb.
4757                  */
4758                 if (PagePrivate(page) &&
4759                     page->private == (unsigned long)eb) {
4760                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4761                         BUG_ON(PageDirty(page));
4762                         BUG_ON(PageWriteback(page));
4763                         /*
4764                          * We need to make sure we haven't be attached
4765                          * to a new eb.
4766                          */
4767                         ClearPagePrivate(page);
4768                         set_page_private(page, 0);
4769                         /* One for the page private */
4770                         put_page(page);
4771                 }
4772
4773                 if (mapped)
4774                         spin_unlock(&page->mapping->private_lock);
4775
4776                 /* One for when we allocated the page */
4777                 put_page(page);
4778         } while (index != 0);
4779 }
4780
4781 /*
4782  * Helper for releasing the extent buffer.
4783  */
4784 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4785 {
4786         btrfs_release_extent_buffer_page(eb);
4787         __free_extent_buffer(eb);
4788 }
4789
4790 static struct extent_buffer *
4791 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4792                       unsigned long len)
4793 {
4794         struct extent_buffer *eb = NULL;
4795
4796         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4797         eb->start = start;
4798         eb->len = len;
4799         eb->fs_info = fs_info;
4800         eb->bflags = 0;
4801         rwlock_init(&eb->lock);
4802         atomic_set(&eb->write_locks, 0);
4803         atomic_set(&eb->read_locks, 0);
4804         atomic_set(&eb->blocking_readers, 0);
4805         atomic_set(&eb->blocking_writers, 0);
4806         atomic_set(&eb->spinning_readers, 0);
4807         atomic_set(&eb->spinning_writers, 0);
4808         eb->lock_nested = 0;
4809         init_waitqueue_head(&eb->write_lock_wq);
4810         init_waitqueue_head(&eb->read_lock_wq);
4811
4812         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4813
4814         spin_lock_init(&eb->refs_lock);
4815         atomic_set(&eb->refs, 1);
4816         atomic_set(&eb->io_pages, 0);
4817
4818         /*
4819          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4820          */
4821         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4822                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4823         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4824
4825         return eb;
4826 }
4827
4828 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4829 {
4830         unsigned long i;
4831         struct page *p;
4832         struct extent_buffer *new;
4833         unsigned long num_pages = num_extent_pages(src->start, src->len);
4834
4835         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4836         if (new == NULL)
4837                 return NULL;
4838
4839         for (i = 0; i < num_pages; i++) {
4840                 p = alloc_page(GFP_NOFS);
4841                 if (!p) {
4842                         btrfs_release_extent_buffer(new);
4843                         return NULL;
4844                 }
4845                 attach_extent_buffer_page(new, p);
4846                 WARN_ON(PageDirty(p));
4847                 SetPageUptodate(p);
4848                 new->pages[i] = p;
4849                 copy_page(page_address(p), page_address(src->pages[i]));
4850         }
4851
4852         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4853         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4854
4855         return new;
4856 }
4857
4858 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4859                                                   u64 start, unsigned long len)
4860 {
4861         struct extent_buffer *eb;
4862         unsigned long num_pages;
4863         unsigned long i;
4864
4865         num_pages = num_extent_pages(start, len);
4866
4867         eb = __alloc_extent_buffer(fs_info, start, len);
4868         if (!eb)
4869                 return NULL;
4870
4871         for (i = 0; i < num_pages; i++) {
4872                 eb->pages[i] = alloc_page(GFP_NOFS);
4873                 if (!eb->pages[i])
4874                         goto err;
4875         }
4876         set_extent_buffer_uptodate(eb);
4877         btrfs_set_header_nritems(eb, 0);
4878         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4879
4880         return eb;
4881 err:
4882         for (; i > 0; i--)
4883                 __free_page(eb->pages[i - 1]);
4884         __free_extent_buffer(eb);
4885         return NULL;
4886 }
4887
4888 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4889                                                 u64 start)
4890 {
4891         return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
4892 }
4893
4894 static void check_buffer_tree_ref(struct extent_buffer *eb)
4895 {
4896         int refs;
4897         /* the ref bit is tricky.  We have to make sure it is set
4898          * if we have the buffer dirty.   Otherwise the
4899          * code to free a buffer can end up dropping a dirty
4900          * page
4901          *
4902          * Once the ref bit is set, it won't go away while the
4903          * buffer is dirty or in writeback, and it also won't
4904          * go away while we have the reference count on the
4905          * eb bumped.
4906          *
4907          * We can't just set the ref bit without bumping the
4908          * ref on the eb because free_extent_buffer might
4909          * see the ref bit and try to clear it.  If this happens
4910          * free_extent_buffer might end up dropping our original
4911          * ref by mistake and freeing the page before we are able
4912          * to add one more ref.
4913          *
4914          * So bump the ref count first, then set the bit.  If someone
4915          * beat us to it, drop the ref we added.
4916          */
4917         refs = atomic_read(&eb->refs);
4918         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4919                 return;
4920
4921         spin_lock(&eb->refs_lock);
4922         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4923                 atomic_inc(&eb->refs);
4924         spin_unlock(&eb->refs_lock);
4925 }
4926
4927 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4928                 struct page *accessed)
4929 {
4930         unsigned long num_pages, i;
4931
4932         check_buffer_tree_ref(eb);
4933
4934         num_pages = num_extent_pages(eb->start, eb->len);
4935         for (i = 0; i < num_pages; i++) {
4936                 struct page *p = eb->pages[i];
4937
4938                 if (p != accessed)
4939                         mark_page_accessed(p);
4940         }
4941 }
4942
4943 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4944                                          u64 start)
4945 {
4946         struct extent_buffer *eb;
4947
4948         rcu_read_lock();
4949         eb = radix_tree_lookup(&fs_info->buffer_radix,
4950                                start >> PAGE_SHIFT);
4951         if (eb && atomic_inc_not_zero(&eb->refs)) {
4952                 rcu_read_unlock();
4953                 /*
4954                  * Lock our eb's refs_lock to avoid races with
4955                  * free_extent_buffer. When we get our eb it might be flagged
4956                  * with EXTENT_BUFFER_STALE and another task running
4957                  * free_extent_buffer might have seen that flag set,
4958                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4959                  * writeback flags not set) and it's still in the tree (flag
4960                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4961                  * of decrementing the extent buffer's reference count twice.
4962                  * So here we could race and increment the eb's reference count,
4963                  * clear its stale flag, mark it as dirty and drop our reference
4964                  * before the other task finishes executing free_extent_buffer,
4965                  * which would later result in an attempt to free an extent
4966                  * buffer that is dirty.
4967                  */
4968                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4969                         spin_lock(&eb->refs_lock);
4970                         spin_unlock(&eb->refs_lock);
4971                 }
4972                 mark_extent_buffer_accessed(eb, NULL);
4973                 return eb;
4974         }
4975         rcu_read_unlock();
4976
4977         return NULL;
4978 }
4979
4980 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4981 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4982                                         u64 start)
4983 {
4984         struct extent_buffer *eb, *exists = NULL;
4985         int ret;
4986
4987         eb = find_extent_buffer(fs_info, start);
4988         if (eb)
4989                 return eb;
4990         eb = alloc_dummy_extent_buffer(fs_info, start);
4991         if (!eb)
4992                 return NULL;
4993         eb->fs_info = fs_info;
4994 again:
4995         ret = radix_tree_preload(GFP_NOFS);
4996         if (ret)
4997                 goto free_eb;
4998         spin_lock(&fs_info->buffer_lock);
4999         ret = radix_tree_insert(&fs_info->buffer_radix,
5000                                 start >> PAGE_SHIFT, eb);
5001         spin_unlock(&fs_info->buffer_lock);
5002         radix_tree_preload_end();
5003         if (ret == -EEXIST) {
5004                 exists = find_extent_buffer(fs_info, start);
5005                 if (exists)
5006                         goto free_eb;
5007                 else
5008                         goto again;
5009         }
5010         check_buffer_tree_ref(eb);
5011         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5012
5013         /*
5014          * We will free dummy extent buffer's if they come into
5015          * free_extent_buffer with a ref count of 2, but if we are using this we
5016          * want the buffers to stay in memory until we're done with them, so
5017          * bump the ref count again.
5018          */
5019         atomic_inc(&eb->refs);
5020         return eb;
5021 free_eb:
5022         btrfs_release_extent_buffer(eb);
5023         return exists;
5024 }
5025 #endif
5026
5027 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
5028                                           u64 start)
5029 {
5030         unsigned long len = fs_info->nodesize;
5031         unsigned long num_pages = num_extent_pages(start, len);
5032         unsigned long i;
5033         unsigned long index = start >> PAGE_SHIFT;
5034         struct extent_buffer *eb;
5035         struct extent_buffer *exists = NULL;
5036         struct page *p;
5037         struct address_space *mapping = fs_info->btree_inode->i_mapping;
5038         int uptodate = 1;
5039         int ret;
5040
5041         if (!IS_ALIGNED(start, fs_info->sectorsize)) {
5042                 btrfs_err(fs_info, "bad tree block start %llu", start);
5043                 return ERR_PTR(-EINVAL);
5044         }
5045
5046         eb = find_extent_buffer(fs_info, start);
5047         if (eb)
5048                 return eb;
5049
5050         eb = __alloc_extent_buffer(fs_info, start, len);
5051         if (!eb)
5052                 return ERR_PTR(-ENOMEM);
5053
5054         for (i = 0; i < num_pages; i++, index++) {
5055                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
5056                 if (!p) {
5057                         exists = ERR_PTR(-ENOMEM);
5058                         goto free_eb;
5059                 }
5060
5061                 spin_lock(&mapping->private_lock);
5062                 if (PagePrivate(p)) {
5063                         /*
5064                          * We could have already allocated an eb for this page
5065                          * and attached one so lets see if we can get a ref on
5066                          * the existing eb, and if we can we know it's good and
5067                          * we can just return that one, else we know we can just
5068                          * overwrite page->private.
5069                          */
5070                         exists = (struct extent_buffer *)p->private;
5071                         if (atomic_inc_not_zero(&exists->refs)) {
5072                                 spin_unlock(&mapping->private_lock);
5073                                 unlock_page(p);
5074                                 put_page(p);
5075                                 mark_extent_buffer_accessed(exists, p);
5076                                 goto free_eb;
5077                         }
5078                         exists = NULL;
5079
5080                         /*
5081                          * Do this so attach doesn't complain and we need to
5082                          * drop the ref the old guy had.
5083                          */
5084                         ClearPagePrivate(p);
5085                         WARN_ON(PageDirty(p));
5086                         put_page(p);
5087                 }
5088                 attach_extent_buffer_page(eb, p);
5089                 spin_unlock(&mapping->private_lock);
5090                 WARN_ON(PageDirty(p));
5091                 eb->pages[i] = p;
5092                 if (!PageUptodate(p))
5093                         uptodate = 0;
5094
5095                 /*
5096                  * see below about how we avoid a nasty race with release page
5097                  * and why we unlock later
5098                  */
5099         }
5100         if (uptodate)
5101                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5102 again:
5103         ret = radix_tree_preload(GFP_NOFS);
5104         if (ret) {
5105                 exists = ERR_PTR(ret);
5106                 goto free_eb;
5107         }
5108
5109         spin_lock(&fs_info->buffer_lock);
5110         ret = radix_tree_insert(&fs_info->buffer_radix,
5111                                 start >> PAGE_SHIFT, eb);
5112         spin_unlock(&fs_info->buffer_lock);
5113         radix_tree_preload_end();
5114         if (ret == -EEXIST) {
5115                 exists = find_extent_buffer(fs_info, start);
5116                 if (exists)
5117                         goto free_eb;
5118                 else
5119                         goto again;
5120         }
5121         /* add one reference for the tree */
5122         check_buffer_tree_ref(eb);
5123         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5124
5125         /*
5126          * there is a race where release page may have
5127          * tried to find this extent buffer in the radix
5128          * but failed.  It will tell the VM it is safe to
5129          * reclaim the, and it will clear the page private bit.
5130          * We must make sure to set the page private bit properly
5131          * after the extent buffer is in the radix tree so
5132          * it doesn't get lost
5133          */
5134         SetPageChecked(eb->pages[0]);
5135         for (i = 1; i < num_pages; i++) {
5136                 p = eb->pages[i];
5137                 ClearPageChecked(p);
5138                 unlock_page(p);
5139         }
5140         unlock_page(eb->pages[0]);
5141         return eb;
5142
5143 free_eb:
5144         WARN_ON(!atomic_dec_and_test(&eb->refs));
5145         for (i = 0; i < num_pages; i++) {
5146                 if (eb->pages[i])
5147                         unlock_page(eb->pages[i]);
5148         }
5149
5150         btrfs_release_extent_buffer(eb);
5151         return exists;
5152 }
5153
5154 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5155 {
5156         struct extent_buffer *eb =
5157                         container_of(head, struct extent_buffer, rcu_head);
5158
5159         __free_extent_buffer(eb);
5160 }
5161
5162 /* Expects to have eb->eb_lock already held */
5163 static int release_extent_buffer(struct extent_buffer *eb)
5164 {
5165         WARN_ON(atomic_read(&eb->refs) == 0);
5166         if (atomic_dec_and_test(&eb->refs)) {
5167                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5168                         struct btrfs_fs_info *fs_info = eb->fs_info;
5169
5170                         spin_unlock(&eb->refs_lock);
5171
5172                         spin_lock(&fs_info->buffer_lock);
5173                         radix_tree_delete(&fs_info->buffer_radix,
5174                                           eb->start >> PAGE_SHIFT);
5175                         spin_unlock(&fs_info->buffer_lock);
5176                 } else {
5177                         spin_unlock(&eb->refs_lock);
5178                 }
5179
5180                 /* Should be safe to release our pages at this point */
5181                 btrfs_release_extent_buffer_page(eb);
5182 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5183                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5184                         __free_extent_buffer(eb);
5185                         return 1;
5186                 }
5187 #endif
5188                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5189                 return 1;
5190         }
5191         spin_unlock(&eb->refs_lock);
5192
5193         return 0;
5194 }
5195
5196 void free_extent_buffer(struct extent_buffer *eb)
5197 {
5198         int refs;
5199         int old;
5200         if (!eb)
5201                 return;
5202
5203         while (1) {
5204                 refs = atomic_read(&eb->refs);
5205                 if (refs <= 3)
5206                         break;
5207                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5208                 if (old == refs)
5209                         return;
5210         }
5211
5212         spin_lock(&eb->refs_lock);
5213         if (atomic_read(&eb->refs) == 2 &&
5214             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5215                 atomic_dec(&eb->refs);
5216
5217         if (atomic_read(&eb->refs) == 2 &&
5218             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5219             !extent_buffer_under_io(eb) &&
5220             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5221                 atomic_dec(&eb->refs);
5222
5223         /*
5224          * I know this is terrible, but it's temporary until we stop tracking
5225          * the uptodate bits and such for the extent buffers.
5226          */
5227         release_extent_buffer(eb);
5228 }
5229
5230 void free_extent_buffer_stale(struct extent_buffer *eb)
5231 {
5232         if (!eb)
5233                 return;
5234
5235         spin_lock(&eb->refs_lock);
5236         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5237
5238         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5239             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5240                 atomic_dec(&eb->refs);
5241         release_extent_buffer(eb);
5242 }
5243
5244 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5245 {
5246         unsigned long i;
5247         unsigned long num_pages;
5248         struct page *page;
5249
5250         num_pages = num_extent_pages(eb->start, eb->len);
5251
5252         for (i = 0; i < num_pages; i++) {
5253                 page = eb->pages[i];
5254                 if (!PageDirty(page))
5255                         continue;
5256
5257                 lock_page(page);
5258                 WARN_ON(!PagePrivate(page));
5259
5260                 clear_page_dirty_for_io(page);
5261                 spin_lock_irq(&page->mapping->tree_lock);
5262                 if (!PageDirty(page)) {
5263                         radix_tree_tag_clear(&page->mapping->page_tree,
5264                                                 page_index(page),
5265                                                 PAGECACHE_TAG_DIRTY);
5266                 }
5267                 spin_unlock_irq(&page->mapping->tree_lock);
5268                 ClearPageError(page);
5269                 unlock_page(page);
5270         }
5271         WARN_ON(atomic_read(&eb->refs) == 0);
5272 }
5273
5274 int set_extent_buffer_dirty(struct extent_buffer *eb)
5275 {
5276         unsigned long i;
5277         unsigned long num_pages;
5278         int was_dirty = 0;
5279
5280         check_buffer_tree_ref(eb);
5281
5282         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5283
5284         num_pages = num_extent_pages(eb->start, eb->len);
5285         WARN_ON(atomic_read(&eb->refs) == 0);
5286         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5287
5288         for (i = 0; i < num_pages; i++)
5289                 set_page_dirty(eb->pages[i]);
5290         return was_dirty;
5291 }
5292
5293 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5294 {
5295         unsigned long i;
5296         struct page *page;
5297         unsigned long num_pages;
5298
5299         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5300         num_pages = num_extent_pages(eb->start, eb->len);
5301         for (i = 0; i < num_pages; i++) {
5302                 page = eb->pages[i];
5303                 if (page)
5304                         ClearPageUptodate(page);
5305         }
5306 }
5307
5308 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5309 {
5310         unsigned long i;
5311         struct page *page;
5312         unsigned long num_pages;
5313
5314         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5315         num_pages = num_extent_pages(eb->start, eb->len);
5316         for (i = 0; i < num_pages; i++) {
5317                 page = eb->pages[i];
5318                 SetPageUptodate(page);
5319         }
5320 }
5321
5322 int extent_buffer_uptodate(struct extent_buffer *eb)
5323 {
5324         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5325 }
5326
5327 int read_extent_buffer_pages(struct extent_io_tree *tree,
5328                              struct extent_buffer *eb, int wait,
5329                              get_extent_t *get_extent, int mirror_num)
5330 {
5331         unsigned long i;
5332         struct page *page;
5333         int err;
5334         int ret = 0;
5335         int locked_pages = 0;
5336         int all_uptodate = 1;
5337         unsigned long num_pages;
5338         unsigned long num_reads = 0;
5339         struct bio *bio = NULL;
5340         unsigned long bio_flags = 0;
5341
5342         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5343                 return 0;
5344
5345         num_pages = num_extent_pages(eb->start, eb->len);
5346         for (i = 0; i < num_pages; i++) {
5347                 page = eb->pages[i];
5348                 if (wait == WAIT_NONE) {
5349                         if (!trylock_page(page))
5350                                 goto unlock_exit;
5351                 } else {
5352                         lock_page(page);
5353                 }
5354                 locked_pages++;
5355         }
5356         /*
5357          * We need to firstly lock all pages to make sure that
5358          * the uptodate bit of our pages won't be affected by
5359          * clear_extent_buffer_uptodate().
5360          */
5361         for (i = 0; i < num_pages; i++) {
5362                 page = eb->pages[i];
5363                 if (!PageUptodate(page)) {
5364                         num_reads++;
5365                         all_uptodate = 0;
5366                 }
5367         }
5368
5369         if (all_uptodate) {
5370                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5371                 goto unlock_exit;
5372         }
5373
5374         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5375         eb->read_mirror = 0;
5376         atomic_set(&eb->io_pages, num_reads);
5377         for (i = 0; i < num_pages; i++) {
5378                 page = eb->pages[i];
5379
5380                 if (!PageUptodate(page)) {
5381                         if (ret) {
5382                                 atomic_dec(&eb->io_pages);
5383                                 unlock_page(page);
5384                                 continue;
5385                         }
5386
5387                         ClearPageError(page);
5388                         err = __extent_read_full_page(tree, page,
5389                                                       get_extent, &bio,
5390                                                       mirror_num, &bio_flags,
5391                                                       REQ_META);
5392                         if (err) {
5393                                 ret = err;
5394                                 /*
5395                                  * We use &bio in above __extent_read_full_page,
5396                                  * so we ensure that if it returns error, the
5397                                  * current page fails to add itself to bio and
5398                                  * it's been unlocked.
5399                                  *
5400                                  * We must dec io_pages by ourselves.
5401                                  */
5402                                 atomic_dec(&eb->io_pages);
5403                         }
5404                 } else {
5405                         unlock_page(page);
5406                 }
5407         }
5408
5409         if (bio) {
5410                 err = submit_one_bio(bio, mirror_num, bio_flags);
5411                 if (err)
5412                         return err;
5413         }
5414
5415         if (ret || wait != WAIT_COMPLETE)
5416                 return ret;
5417
5418         for (i = 0; i < num_pages; i++) {
5419                 page = eb->pages[i];
5420                 wait_on_page_locked(page);
5421                 if (!PageUptodate(page))
5422                         ret = -EIO;
5423         }
5424
5425         return ret;
5426
5427 unlock_exit:
5428         while (locked_pages > 0) {
5429                 locked_pages--;
5430                 page = eb->pages[locked_pages];
5431                 unlock_page(page);
5432         }
5433         return ret;
5434 }
5435
5436 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
5437                         unsigned long start,
5438                         unsigned long len)
5439 {
5440         size_t cur;
5441         size_t offset;
5442         struct page *page;
5443         char *kaddr;
5444         char *dst = (char *)dstv;
5445         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5446         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5447
5448         WARN_ON(start > eb->len);
5449         WARN_ON(start + len > eb->start + eb->len);
5450
5451         offset = (start_offset + start) & (PAGE_SIZE - 1);
5452
5453         while (len > 0) {
5454                 page = eb->pages[i];
5455
5456                 cur = min(len, (PAGE_SIZE - offset));
5457                 kaddr = page_address(page);
5458                 memcpy(dst, kaddr + offset, cur);
5459
5460                 dst += cur;
5461                 len -= cur;
5462                 offset = 0;
5463                 i++;
5464         }
5465 }
5466
5467 int read_extent_buffer_to_user(struct extent_buffer *eb, void __user *dstv,
5468                         unsigned long start,
5469                         unsigned long len)
5470 {
5471         size_t cur;
5472         size_t offset;
5473         struct page *page;
5474         char *kaddr;
5475         char __user *dst = (char __user *)dstv;
5476         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5477         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5478         int ret = 0;
5479
5480         WARN_ON(start > eb->len);
5481         WARN_ON(start + len > eb->start + eb->len);
5482
5483         offset = (start_offset + start) & (PAGE_SIZE - 1);
5484
5485         while (len > 0) {
5486                 page = eb->pages[i];
5487
5488                 cur = min(len, (PAGE_SIZE - offset));
5489                 kaddr = page_address(page);
5490                 if (copy_to_user(dst, kaddr + offset, cur)) {
5491                         ret = -EFAULT;
5492                         break;
5493                 }
5494
5495                 dst += cur;
5496                 len -= cur;
5497                 offset = 0;
5498                 i++;
5499         }
5500
5501         return ret;
5502 }
5503
5504 /*
5505  * return 0 if the item is found within a page.
5506  * return 1 if the item spans two pages.
5507  * return -EINVAL otherwise.
5508  */
5509 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
5510                                unsigned long min_len, char **map,
5511                                unsigned long *map_start,
5512                                unsigned long *map_len)
5513 {
5514         size_t offset = start & (PAGE_SIZE - 1);
5515         char *kaddr;
5516         struct page *p;
5517         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5518         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5519         unsigned long end_i = (start_offset + start + min_len - 1) >>
5520                 PAGE_SHIFT;
5521
5522         if (i != end_i)
5523                 return 1;
5524
5525         if (i == 0) {
5526                 offset = start_offset;
5527                 *map_start = 0;
5528         } else {
5529                 offset = 0;
5530                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5531         }
5532
5533         if (start + min_len > eb->len) {
5534                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5535                        eb->start, eb->len, start, min_len);
5536                 return -EINVAL;
5537         }
5538
5539         p = eb->pages[i];
5540         kaddr = page_address(p);
5541         *map = kaddr + offset;
5542         *map_len = PAGE_SIZE - offset;
5543         return 0;
5544 }
5545
5546 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
5547                           unsigned long start,
5548                           unsigned long len)
5549 {
5550         size_t cur;
5551         size_t offset;
5552         struct page *page;
5553         char *kaddr;
5554         char *ptr = (char *)ptrv;
5555         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5556         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5557         int ret = 0;
5558
5559         WARN_ON(start > eb->len);
5560         WARN_ON(start + len > eb->start + eb->len);
5561
5562         offset = (start_offset + start) & (PAGE_SIZE - 1);
5563
5564         while (len > 0) {
5565                 page = eb->pages[i];
5566
5567                 cur = min(len, (PAGE_SIZE - offset));
5568
5569                 kaddr = page_address(page);
5570                 ret = memcmp(ptr, kaddr + offset, cur);
5571                 if (ret)
5572                         break;
5573
5574                 ptr += cur;
5575                 len -= cur;
5576                 offset = 0;
5577                 i++;
5578         }
5579         return ret;
5580 }
5581
5582 void write_extent_buffer_chunk_tree_uuid(struct extent_buffer *eb,
5583                 const void *srcv)
5584 {
5585         char *kaddr;
5586
5587         WARN_ON(!PageUptodate(eb->pages[0]));
5588         kaddr = page_address(eb->pages[0]);
5589         memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv,
5590                         BTRFS_FSID_SIZE);
5591 }
5592
5593 void write_extent_buffer_fsid(struct extent_buffer *eb, const void *srcv)
5594 {
5595         char *kaddr;
5596
5597         WARN_ON(!PageUptodate(eb->pages[0]));
5598         kaddr = page_address(eb->pages[0]);
5599         memcpy(kaddr + offsetof(struct btrfs_header, fsid), srcv,
5600                         BTRFS_FSID_SIZE);
5601 }
5602
5603 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5604                          unsigned long start, unsigned long len)
5605 {
5606         size_t cur;
5607         size_t offset;
5608         struct page *page;
5609         char *kaddr;
5610         char *src = (char *)srcv;
5611         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5612         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5613
5614         WARN_ON(start > eb->len);
5615         WARN_ON(start + len > eb->start + eb->len);
5616
5617         offset = (start_offset + start) & (PAGE_SIZE - 1);
5618
5619         while (len > 0) {
5620                 page = eb->pages[i];
5621                 WARN_ON(!PageUptodate(page));
5622
5623                 cur = min(len, PAGE_SIZE - offset);
5624                 kaddr = page_address(page);
5625                 memcpy(kaddr + offset, src, cur);
5626
5627                 src += cur;
5628                 len -= cur;
5629                 offset = 0;
5630                 i++;
5631         }
5632 }
5633
5634 void memzero_extent_buffer(struct extent_buffer *eb, unsigned long start,
5635                 unsigned long len)
5636 {
5637         size_t cur;
5638         size_t offset;
5639         struct page *page;
5640         char *kaddr;
5641         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5642         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5643
5644         WARN_ON(start > eb->len);
5645         WARN_ON(start + len > eb->start + eb->len);
5646
5647         offset = (start_offset + start) & (PAGE_SIZE - 1);
5648
5649         while (len > 0) {
5650                 page = eb->pages[i];
5651                 WARN_ON(!PageUptodate(page));
5652
5653                 cur = min(len, PAGE_SIZE - offset);
5654                 kaddr = page_address(page);
5655                 memset(kaddr + offset, 0, cur);
5656
5657                 len -= cur;
5658                 offset = 0;
5659                 i++;
5660         }
5661 }
5662
5663 void copy_extent_buffer_full(struct extent_buffer *dst,
5664                              struct extent_buffer *src)
5665 {
5666         int i;
5667         unsigned num_pages;
5668
5669         ASSERT(dst->len == src->len);
5670
5671         num_pages = num_extent_pages(dst->start, dst->len);
5672         for (i = 0; i < num_pages; i++)
5673                 copy_page(page_address(dst->pages[i]),
5674                                 page_address(src->pages[i]));
5675 }
5676
5677 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5678                         unsigned long dst_offset, unsigned long src_offset,
5679                         unsigned long len)
5680 {
5681         u64 dst_len = dst->len;
5682         size_t cur;
5683         size_t offset;
5684         struct page *page;
5685         char *kaddr;
5686         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5687         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5688
5689         WARN_ON(src->len != dst_len);
5690
5691         offset = (start_offset + dst_offset) &
5692                 (PAGE_SIZE - 1);
5693
5694         while (len > 0) {
5695                 page = dst->pages[i];
5696                 WARN_ON(!PageUptodate(page));
5697
5698                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5699
5700                 kaddr = page_address(page);
5701                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5702
5703                 src_offset += cur;
5704                 len -= cur;
5705                 offset = 0;
5706                 i++;
5707         }
5708 }
5709
5710 void le_bitmap_set(u8 *map, unsigned int start, int len)
5711 {
5712         u8 *p = map + BIT_BYTE(start);
5713         const unsigned int size = start + len;
5714         int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5715         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);
5716
5717         while (len - bits_to_set >= 0) {
5718                 *p |= mask_to_set;
5719                 len -= bits_to_set;
5720                 bits_to_set = BITS_PER_BYTE;
5721                 mask_to_set = ~0;
5722                 p++;
5723         }
5724         if (len) {
5725                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5726                 *p |= mask_to_set;
5727         }
5728 }
5729
5730 void le_bitmap_clear(u8 *map, unsigned int start, int len)
5731 {
5732         u8 *p = map + BIT_BYTE(start);
5733         const unsigned int size = start + len;
5734         int bits_to_clear = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5735         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(start);
5736
5737         while (len - bits_to_clear >= 0) {
5738                 *p &= ~mask_to_clear;
5739                 len -= bits_to_clear;
5740                 bits_to_clear = BITS_PER_BYTE;
5741                 mask_to_clear = ~0;
5742                 p++;
5743         }
5744         if (len) {
5745                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5746                 *p &= ~mask_to_clear;
5747         }
5748 }
5749
5750 /*
5751  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5752  * given bit number
5753  * @eb: the extent buffer
5754  * @start: offset of the bitmap item in the extent buffer
5755  * @nr: bit number
5756  * @page_index: return index of the page in the extent buffer that contains the
5757  * given bit number
5758  * @page_offset: return offset into the page given by page_index
5759  *
5760  * This helper hides the ugliness of finding the byte in an extent buffer which
5761  * contains a given bit.
5762  */
5763 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5764                                     unsigned long start, unsigned long nr,
5765                                     unsigned long *page_index,
5766                                     size_t *page_offset)
5767 {
5768         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5769         size_t byte_offset = BIT_BYTE(nr);
5770         size_t offset;
5771
5772         /*
5773          * The byte we want is the offset of the extent buffer + the offset of
5774          * the bitmap item in the extent buffer + the offset of the byte in the
5775          * bitmap item.
5776          */
5777         offset = start_offset + start + byte_offset;
5778
5779         *page_index = offset >> PAGE_SHIFT;
5780         *page_offset = offset & (PAGE_SIZE - 1);
5781 }
5782
5783 /**
5784  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5785  * @eb: the extent buffer
5786  * @start: offset of the bitmap item in the extent buffer
5787  * @nr: bit number to test
5788  */
5789 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5790                            unsigned long nr)
5791 {
5792         u8 *kaddr;
5793         struct page *page;
5794         unsigned long i;
5795         size_t offset;
5796
5797         eb_bitmap_offset(eb, start, nr, &i, &offset);
5798         page = eb->pages[i];
5799         WARN_ON(!PageUptodate(page));
5800         kaddr = page_address(page);
5801         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5802 }
5803
5804 /**
5805  * extent_buffer_bitmap_set - set an area of a bitmap
5806  * @eb: the extent buffer
5807  * @start: offset of the bitmap item in the extent buffer
5808  * @pos: bit number of the first bit
5809  * @len: number of bits to set
5810  */
5811 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5812                               unsigned long pos, unsigned long len)
5813 {
5814         u8 *kaddr;
5815         struct page *page;
5816         unsigned long i;
5817         size_t offset;
5818         const unsigned int size = pos + len;
5819         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5820         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5821
5822         eb_bitmap_offset(eb, start, pos, &i, &offset);
5823         page = eb->pages[i];
5824         WARN_ON(!PageUptodate(page));
5825         kaddr = page_address(page);
5826
5827         while (len >= bits_to_set) {
5828                 kaddr[offset] |= mask_to_set;
5829                 len -= bits_to_set;
5830                 bits_to_set = BITS_PER_BYTE;
5831                 mask_to_set = ~0;
5832                 if (++offset >= PAGE_SIZE && len > 0) {
5833                         offset = 0;
5834                         page = eb->pages[++i];
5835                         WARN_ON(!PageUptodate(page));
5836                         kaddr = page_address(page);
5837                 }
5838         }
5839         if (len) {
5840                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5841                 kaddr[offset] |= mask_to_set;
5842         }
5843 }
5844
5845
5846 /**
5847  * extent_buffer_bitmap_clear - clear an area of a bitmap
5848  * @eb: the extent buffer
5849  * @start: offset of the bitmap item in the extent buffer
5850  * @pos: bit number of the first bit
5851  * @len: number of bits to clear
5852  */
5853 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5854                                 unsigned long pos, unsigned long len)
5855 {
5856         u8 *kaddr;
5857         struct page *page;
5858         unsigned long i;
5859         size_t offset;
5860         const unsigned int size = pos + len;
5861         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5862         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5863
5864         eb_bitmap_offset(eb, start, pos, &i, &offset);
5865         page = eb->pages[i];
5866         WARN_ON(!PageUptodate(page));
5867         kaddr = page_address(page);
5868
5869         while (len >= bits_to_clear) {
5870                 kaddr[offset] &= ~mask_to_clear;
5871                 len -= bits_to_clear;
5872                 bits_to_clear = BITS_PER_BYTE;
5873                 mask_to_clear = ~0;
5874                 if (++offset >= PAGE_SIZE && len > 0) {
5875                         offset = 0;
5876                         page = eb->pages[++i];
5877                         WARN_ON(!PageUptodate(page));
5878                         kaddr = page_address(page);
5879                 }
5880         }
5881         if (len) {
5882                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5883                 kaddr[offset] &= ~mask_to_clear;
5884         }
5885 }
5886
5887 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5888 {
5889         unsigned long distance = (src > dst) ? src - dst : dst - src;
5890         return distance < len;
5891 }
5892
5893 static void copy_pages(struct page *dst_page, struct page *src_page,
5894                        unsigned long dst_off, unsigned long src_off,
5895                        unsigned long len)
5896 {
5897         char *dst_kaddr = page_address(dst_page);
5898         char *src_kaddr;
5899         int must_memmove = 0;
5900
5901         if (dst_page != src_page) {
5902                 src_kaddr = page_address(src_page);
5903         } else {
5904                 src_kaddr = dst_kaddr;
5905                 if (areas_overlap(src_off, dst_off, len))
5906                         must_memmove = 1;
5907         }
5908
5909         if (must_memmove)
5910                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5911         else
5912                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5913 }
5914
5915 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5916                            unsigned long src_offset, unsigned long len)
5917 {
5918         struct btrfs_fs_info *fs_info = dst->fs_info;
5919         size_t cur;
5920         size_t dst_off_in_page;
5921         size_t src_off_in_page;
5922         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5923         unsigned long dst_i;
5924         unsigned long src_i;
5925
5926         if (src_offset + len > dst->len) {
5927                 btrfs_err(fs_info,
5928                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5929                          src_offset, len, dst->len);
5930                 BUG_ON(1);
5931         }
5932         if (dst_offset + len > dst->len) {
5933                 btrfs_err(fs_info,
5934                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5935                          dst_offset, len, dst->len);
5936                 BUG_ON(1);
5937         }
5938
5939         while (len > 0) {
5940                 dst_off_in_page = (start_offset + dst_offset) &
5941                         (PAGE_SIZE - 1);
5942                 src_off_in_page = (start_offset + src_offset) &
5943                         (PAGE_SIZE - 1);
5944
5945                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5946                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5947
5948                 cur = min(len, (unsigned long)(PAGE_SIZE -
5949                                                src_off_in_page));
5950                 cur = min_t(unsigned long, cur,
5951                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5952
5953                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5954                            dst_off_in_page, src_off_in_page, cur);
5955
5956                 src_offset += cur;
5957                 dst_offset += cur;
5958                 len -= cur;
5959         }
5960 }
5961
5962 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5963                            unsigned long src_offset, unsigned long len)
5964 {
5965         struct btrfs_fs_info *fs_info = dst->fs_info;
5966         size_t cur;
5967         size_t dst_off_in_page;
5968         size_t src_off_in_page;
5969         unsigned long dst_end = dst_offset + len - 1;
5970         unsigned long src_end = src_offset + len - 1;
5971         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5972         unsigned long dst_i;
5973         unsigned long src_i;
5974
5975         if (src_offset + len > dst->len) {
5976                 btrfs_err(fs_info,
5977                           "memmove bogus src_offset %lu move len %lu len %lu",
5978                           src_offset, len, dst->len);
5979                 BUG_ON(1);
5980         }
5981         if (dst_offset + len > dst->len) {
5982                 btrfs_err(fs_info,
5983                           "memmove bogus dst_offset %lu move len %lu len %lu",
5984                           dst_offset, len, dst->len);
5985                 BUG_ON(1);
5986         }
5987         if (dst_offset < src_offset) {
5988                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5989                 return;
5990         }
5991         while (len > 0) {
5992                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5993                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5994
5995                 dst_off_in_page = (start_offset + dst_end) &
5996                         (PAGE_SIZE - 1);
5997                 src_off_in_page = (start_offset + src_end) &
5998                         (PAGE_SIZE - 1);
5999
6000                 cur = min_t(unsigned long, len, src_off_in_page + 1);
6001                 cur = min(cur, dst_off_in_page + 1);
6002                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
6003                            dst_off_in_page - cur + 1,
6004                            src_off_in_page - cur + 1, cur);
6005
6006                 dst_end -= cur;
6007                 src_end -= cur;
6008                 len -= cur;
6009         }
6010 }
6011
6012 int try_release_extent_buffer(struct page *page)
6013 {
6014         struct extent_buffer *eb;
6015
6016         /*
6017          * We need to make sure nobody is attaching this page to an eb right
6018          * now.
6019          */
6020         spin_lock(&page->mapping->private_lock);
6021         if (!PagePrivate(page)) {
6022                 spin_unlock(&page->mapping->private_lock);
6023                 return 1;
6024         }
6025
6026         eb = (struct extent_buffer *)page->private;
6027         BUG_ON(!eb);
6028
6029         /*
6030          * This is a little awful but should be ok, we need to make sure that
6031          * the eb doesn't disappear out from under us while we're looking at
6032          * this page.
6033          */
6034         spin_lock(&eb->refs_lock);
6035         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
6036                 spin_unlock(&eb->refs_lock);
6037                 spin_unlock(&page->mapping->private_lock);
6038                 return 0;
6039         }
6040         spin_unlock(&page->mapping->private_lock);
6041
6042         /*
6043          * If tree ref isn't set then we know the ref on this eb is a real ref,
6044          * so just return, this page will likely be freed soon anyway.
6045          */
6046         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
6047                 spin_unlock(&eb->refs_lock);
6048                 return 0;
6049         }
6050
6051         return release_extent_buffer(eb);
6052 }